ayrisdev/crawl4ai
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
c0fd36982d26099efeb90e52600500d032122384
crawl4ai/docs/apps/iseeyou/llms-full.txt

7716 lines
237 KiB
Plaintext
Raw Blame History

This file contains invisible Unicode characters
This file contains invisible Unicode characters that are indistinguishable to humans but may be processed differently by a computer. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# Crawl4AI
> Open-source LLM-friendly web crawler and scraper for AI applications
Crawl4AI simplifies asynchronous web crawling and data extraction, making it accessible for large language models (LLMs) and AI applications. Built with Python and Playwright for high-performance crawling with structured data extraction.
**Key Features:**
- Asynchronous crawling with high concurrency
- Multiple extraction strategies (CSS, XPath, LLM-based)
- Built-in markdown generation with content filtering
- Docker deployment with REST API
- Session management and browser automation
- Advanced anti-detection capabilities
**Quick Links:**
- [GitHub Repository](https://github.com/unclecode/crawl4ai)
- [Documentation](https://docs.crawl4ai.com)
- [Examples](https://github.com/unclecode/crawl4ai/tree/main/docs/examples)
---
## Installation
Multiple installation options for different environments and use cases.
### Basic Installation
```bash
# Install core library
pip install crawl4ai
# Initial setup (installs Playwright browsers)
crawl4ai-setup
# Verify installation
crawl4ai-doctor
```
### Quick Verification
```python
import asyncio
from crawl4ai import AsyncWebCrawler
async def main():
async with AsyncWebCrawler() as crawler:
result = await crawler.arun("https://example.com")
print(result.markdown[:300])
if __name__ == "__main__":
asyncio.run(main())
```
**📖 Learn more:** [Basic Usage Guide](https://docs.crawl4ai.com/core/quickstart.md)
### Advanced Features (Optional)
```bash
# PyTorch-based features (text clustering, semantic chunking)
pip install crawl4ai[torch]
crawl4ai-setup
# Transformers (Hugging Face models)
pip install crawl4ai[transformer]
crawl4ai-setup
# All features (large download)
pip install crawl4ai[all]
crawl4ai-setup
# Pre-download models (optional)
crawl4ai-download-models
```
**📖 Learn more:** [Advanced Features Documentation](https://docs.crawl4ai.com/extraction/llm-strategies.md)
### Docker Deployment
```bash
# Pull pre-built image (specify platform for consistency)
docker pull --platform linux/amd64 unclecode/crawl4ai:latest
# For ARM (M1/M2 Macs): docker pull --platform linux/arm64 unclecode/crawl4ai:latest
# Setup environment for LLM support
cat > .llm.env << EOL
OPENAI_API_KEY=sk-your-key
ANTHROPIC_API_KEY=your-anthropic-key
EOL
# Run with LLM support (specify platform)
docker run -d \
--platform linux/amd64 \
-p 11235:11235 \
--name crawl4ai \
--env-file .llm.env \
--shm-size=1g \
unclecode/crawl4ai:latest
# For ARM Macs, use: --platform linux/arm64
# Basic run (no LLM)
docker run -d \
--platform linux/amd64 \
-p 11235:11235 \
--name crawl4ai \
--shm-size=1g \
unclecode/crawl4ai:latest
```
**📖 Learn more:** [Complete Docker Guide](https://docs.crawl4ai.com/core/docker-deployment.md)
### Docker Compose
```bash
# Clone repository
git clone https://github.com/unclecode/crawl4ai.git
cd crawl4ai
# Copy environment template
cp deploy/docker/.llm.env.example .llm.env
# Edit .llm.env with your API keys
# Run pre-built image
IMAGE=unclecode/crawl4ai:latest docker compose up -d
# Build and run locally
docker compose up --build -d
# Build with all features
INSTALL_TYPE=all docker compose up --build -d
# Stop service
docker compose down
```
**📖 Learn more:** [Docker Compose Configuration](https://docs.crawl4ai.com/core/docker-deployment.md#option-2-using-docker-compose)
### Manual Docker Build
```bash
# Build multi-architecture image (specify platform)
docker buildx build --platform linux/amd64 -t crawl4ai-local:latest --load .
# For ARM: docker buildx build --platform linux/arm64 -t crawl4ai-local:latest --load .
# Build with specific features
docker buildx build \
--platform linux/amd64 \
--build-arg INSTALL_TYPE=all \
--build-arg ENABLE_GPU=false \
-t crawl4ai-local:latest --load .
# Run custom build (specify platform)
docker run -d \
--platform linux/amd64 \
-p 11235:11235 \
--name crawl4ai-custom \
--env-file .llm.env \
--shm-size=1g \
crawl4ai-local:latest
```
**📖 Learn more:** [Manual Build Guide](https://docs.crawl4ai.com/core/docker-deployment.md#option-3-manual-local-build--run)
### Google Colab
```python
# Install in Colab
!pip install crawl4ai
!crawl4ai-setup
# If setup fails, manually install Playwright browsers
!playwright install chromium
# Install with all features (may take 5-10 minutes)
!pip install crawl4ai[all]
!crawl4ai-setup
!crawl4ai-download-models
# If still having issues, force Playwright install
!playwright install chromium --force
# Quick test
import asyncio
from crawl4ai import AsyncWebCrawler
async def test_crawl():
async with AsyncWebCrawler() as crawler:
result = await crawler.arun("https://example.com")
print("✅ Installation successful!")
print(f"Content length: {len(result.markdown)}")
# Run test in Colab
await test_crawl()
```
**📖 Learn more:** [Colab Examples Notebook](https://colab.research.google.com/github/unclecode/crawl4ai/blob/main/docs/examples/quickstart.ipynb)
### Docker API Usage
```python
# Using Docker SDK
import asyncio
from crawl4ai.docker_client import Crawl4aiDockerClient
from crawl4ai import BrowserConfig, CrawlerRunConfig, CacheMode
async def main():
async with Crawl4aiDockerClient(base_url="http://localhost:11235") as client:
results = await client.crawl(
["https://example.com"],
browser_config=BrowserConfig(headless=True),
crawler_config=CrawlerRunConfig(cache_mode=CacheMode.BYPASS)
)
for result in results:
print(f"Success: {result.success}, Length: {len(result.markdown)}")
asyncio.run(main())
```
**📖 Learn more:** [Docker Client API](https://docs.crawl4ai.com/core/docker-deployment.md#python-sdk)
### Direct API Calls
```python
# REST API example
import requests
payload = {
"urls": ["https://example.com"],
"browser_config": {"type": "BrowserConfig", "params": {"headless": True}},
"crawler_config": {"type": "CrawlerRunConfig", "params": {"cache_mode": "bypass"}}
}
response = requests.post("http://localhost:11235/crawl", json=payload)
print(response.json())
```
**📖 Learn more:** [REST API Reference](https://docs.crawl4ai.com/core/docker-deployment.md#rest-api-examples)
### Health Check
```bash
# Check Docker service
curl http://localhost:11235/health
# Access playground
open http://localhost:11235/playground
# View metrics
curl http://localhost:11235/metrics
```
**📖 Learn more:** [Monitoring & Metrics](https://docs.crawl4ai.com/core/docker-deployment.md#metrics--monitoring)
---
## Simple Crawling
Basic web crawling operations with AsyncWebCrawler, configurations, and response handling.
### Basic Setup
```python
import asyncio
from crawl4ai import AsyncWebCrawler, BrowserConfig, CrawlerRunConfig
async def main():
browser_config = BrowserConfig() # Default browser settings
run_config = CrawlerRunConfig() # Default crawl settings
async with AsyncWebCrawler(config=browser_config) as crawler:
result = await crawler.arun(
url="https://example.com",
config=run_config
)
print(result.markdown)
if __name__ == "__main__":
asyncio.run(main())
```
### Understanding CrawlResult
```python
from crawl4ai.markdown_generation_strategy import DefaultMarkdownGenerator
from crawl4ai.content_filter_strategy import PruningContentFilter
config = CrawlerRunConfig(
markdown_generator=DefaultMarkdownGenerator(
content_filter=PruningContentFilter(threshold=0.6),
options={"ignore_links": True}
)
)
result = await crawler.arun("https://example.com", config=config)
# Different content formats
print(result.html) # Raw HTML
print(result.cleaned_html) # Cleaned HTML
print(result.markdown.raw_markdown) # Raw markdown
print(result.markdown.fit_markdown) # Filtered markdown
# Status information
print(result.success) # True/False
print(result.status_code) # HTTP status (200, 404, etc.)
# Extracted content
print(result.media) # Images, videos, audio
print(result.links) # Internal/external links
```
### Basic Configuration Options
```python
run_config = CrawlerRunConfig(
word_count_threshold=10, # Min words per block
exclude_external_links=True, # Remove external links
remove_overlay_elements=True, # Remove popups/modals
process_iframes=True, # Process iframe content
excluded_tags=['form', 'header'] # Skip these tags
)
result = await crawler.arun("https://example.com", config=run_config)
```
### Error Handling
```python
result = await crawler.arun("https://example.com", config=run_config)
if not result.success:
print(f"Crawl failed: {result.error_message}")
print(f"Status code: {result.status_code}")
else:
print(f"Success! Content length: {len(result.markdown)}")
```
### Debugging with Verbose Logging
```python
browser_config = BrowserConfig(verbose=True)
async with AsyncWebCrawler(config=browser_config) as crawler:
result = await crawler.arun("https://example.com")
# Detailed logging output will be displayed
```
### Complete Example
```python
import asyncio
from crawl4ai import AsyncWebCrawler, BrowserConfig, CrawlerRunConfig, CacheMode
async def comprehensive_crawl():
browser_config = BrowserConfig(verbose=True)
run_config = CrawlerRunConfig(
# Content filtering
word_count_threshold=10,
excluded_tags=['form', 'header', 'nav'],
exclude_external_links=True,
# Content processing
process_iframes=True,
remove_overlay_elements=True,
# Cache control
cache_mode=CacheMode.ENABLED
)
async with AsyncWebCrawler(config=browser_config) as crawler:
result = await crawler.arun(
url="https://example.com",
config=run_config
)
if result.success:
# Display content summary
print(f"Title: {result.metadata.get('title', 'No title')}")
print(f"Content: {result.markdown[:500]}...")
# Process media
images = result.media.get("images", [])
print(f"Found {len(images)} images")
for img in images[:3]: # First 3 images
print(f" - {img.get('src', 'No src')}")
# Process links
internal_links = result.links.get("internal", [])
print(f"Found {len(internal_links)} internal links")
for link in internal_links[:3]: # First 3 links
print(f" - {link.get('href', 'No href')}")
else:
print(f"❌ Crawl failed: {result.error_message}")
print(f"Status: {result.status_code}")
if __name__ == "__main__":
asyncio.run(comprehensive_crawl())
```
### Working with Raw HTML and Local Files
```python
# Crawl raw HTML
raw_html = "<html><body><h1>Test</h1><p>Content</p></body></html>"
result = await crawler.arun(f"raw://{raw_html}")
# Crawl local file
result = await crawler.arun("file:///path/to/local/file.html")
# Both return standard CrawlResult objects
print(result.markdown)
```
## Table Extraction
Extract structured data from HTML tables with automatic detection and scoring.
### Basic Table Extraction
```python
import asyncio
import pandas as pd
from crawl4ai import AsyncWebCrawler, CrawlerRunConfig, CacheMode
async def extract_tables():
async with AsyncWebCrawler() as crawler:
config = CrawlerRunConfig(
table_score_threshold=7, # Higher = stricter detection
cache_mode=CacheMode.BYPASS
)
result = await crawler.arun("https://example.com/tables", config=config)
if result.success and result.tables:
# New tables field (v0.6+)
for i, table in enumerate(result.tables):
print(f"Table {i+1}:")
print(f"Headers: {table['headers']}")
print(f"Rows: {len(table['rows'])}")
print(f"Caption: {table.get('caption', 'No caption')}")
# Convert to DataFrame
df = pd.DataFrame(table['rows'], columns=table['headers'])
print(df.head())
asyncio.run(extract_tables())
```
### Advanced Table Processing
```python
from crawl4ai import LXMLWebScrapingStrategy
async def process_financial_tables():
config = CrawlerRunConfig(
table_score_threshold=8, # Strict detection for data tables
scraping_strategy=LXMLWebScrapingStrategy(),
keep_data_attributes=True,
scan_full_page=True
)
async with AsyncWebCrawler() as crawler:
result = await crawler.arun("https://coinmarketcap.com", config=config)
if result.tables:
# Get the main data table (usually first/largest)
main_table = result.tables[0]
# Create DataFrame
df = pd.DataFrame(
main_table['rows'],
columns=main_table['headers']
)
# Clean and process data
df = clean_financial_data(df)
# Save for analysis
df.to_csv("market_data.csv", index=False)
return df
def clean_financial_data(df):
"""Clean currency symbols, percentages, and large numbers"""
for col in df.columns:
if 'price' in col.lower():
# Remove currency symbols
df[col] = df[col].str.replace(r'[^\d.]', '', regex=True)
df[col] = pd.to_numeric(df[col], errors='coerce')
elif '%' in str(df[col].iloc[0]):
# Convert percentages
df[col] = df[col].str.replace('%', '').astype(float) / 100
elif any(suffix in str(df[col].iloc[0]) for suffix in ['B', 'M', 'K']):
# Handle large numbers (Billions, Millions, etc.)
df[col] = df[col].apply(convert_large_numbers)
return df
def convert_large_numbers(value):
"""Convert 1.5B -> 1500000000"""
if pd.isna(value):
return float('nan')
value = str(value)
multiplier = 1
if 'B' in value:
multiplier = 1e9
elif 'M' in value:
multiplier = 1e6
elif 'K' in value:
multiplier = 1e3
number = float(re.sub(r'[^\d.]', '', value))
return number * multiplier
```
### Table Detection Configuration
```python
# Strict table detection (data-heavy pages)
strict_config = CrawlerRunConfig(
table_score_threshold=9, # Only high-quality tables
word_count_threshold=5, # Ignore sparse content
excluded_tags=['nav', 'footer'] # Skip navigation tables
)
# Lenient detection (mixed content pages)
lenient_config = CrawlerRunConfig(
table_score_threshold=5, # Include layout tables
process_iframes=True, # Check embedded tables
scan_full_page=True # Scroll to load dynamic tables
)
# Financial/data site optimization
financial_config = CrawlerRunConfig(
table_score_threshold=8,
scraping_strategy=LXMLWebScrapingStrategy(),
wait_for="css:table", # Wait for tables to load
scan_full_page=True,
scroll_delay=0.2
)
```
### Multi-Table Processing
```python
async def extract_all_tables():
async with AsyncWebCrawler() as crawler:
result = await crawler.arun("https://example.com/data", config=config)
tables_data = {}
for i, table in enumerate(result.tables):
# Create meaningful names based on content
table_name = (
table.get('caption') or
f"table_{i+1}_{table['headers'][0]}"
).replace(' ', '_').lower()
df = pd.DataFrame(table['rows'], columns=table['headers'])
# Store with metadata
tables_data[table_name] = {
'dataframe': df,
'headers': table['headers'],
'row_count': len(table['rows']),
'caption': table.get('caption'),
'summary': table.get('summary')
}
return tables_data
# Usage
tables = await extract_all_tables()
for name, data in tables.items():
print(f"{name}: {data['row_count']} rows")
data['dataframe'].to_csv(f"{name}.csv")
```
### Backward Compatibility
```python
# Support both new and old table formats
def get_tables(result):
# New format (v0.6+)
if hasattr(result, 'tables') and result.tables:
return result.tables
# Fallback to media.tables (older versions)
return result.media.get('tables', [])
# Usage in existing code
result = await crawler.arun(url, config=config)
tables = get_tables(result)
for table in tables:
df = pd.DataFrame(table['rows'], columns=table['headers'])
# Process table data...
```
### Table Quality Scoring
```python
# Understanding table_score_threshold values:
# 10: Only perfect data tables (headers + data rows)
# 8-9: High-quality tables (recommended for financial/data sites)
# 6-7: Mixed content tables (news sites, wikis)
# 4-5: Layout tables included (broader detection)
# 1-3: All table-like structures (very permissive)
config = CrawlerRunConfig(
table_score_threshold=8, # Balanced detection
verbose=True # See scoring details in logs
)
```
**📖 Learn more:** [CrawlResult API Reference](https://docs.crawl4ai.com/api/crawl-result/), [Browser & Crawler Configuration](https://docs.crawl4ai.com/core/browser-crawler-config/), [Cache Modes](https://docs.crawl4ai.com/core/cache-modes/)
---
## Browser, Crawler & LLM Configuration
Core configuration classes for controlling browser behavior, crawl operations, LLM providers, and understanding crawl results.
### BrowserConfig - Browser Environment Setup
```python
from crawl4ai import BrowserConfig, AsyncWebCrawler
# Basic browser configuration
browser_config = BrowserConfig(
browser_type="chromium", # "chromium", "firefox", "webkit"
headless=True, # False for visible browser (debugging)
viewport_width=1280,
viewport_height=720,
verbose=True
)
# Advanced browser setup with proxy and persistence
browser_config = BrowserConfig(
headless=False,
proxy="http://user:pass@proxy:8080",
use_persistent_context=True,
user_data_dir="./browser_data",
cookies=[
{"name": "session", "value": "abc123", "domain": "example.com"}
],
headers={"Accept-Language": "en-US,en;q=0.9"},
user_agent="Mozilla/5.0 (X11; Linux x86_64) Chrome/116.0.0.0 Safari/537.36",
text_mode=True, # Disable images for faster crawling
extra_args=["--disable-extensions", "--no-sandbox"]
)
async with AsyncWebCrawler(config=browser_config) as crawler:
result = await crawler.arun("https://example.com")
```
### CrawlerRunConfig - Crawl Operation Control
```python
from crawl4ai import CrawlerRunConfig, CacheMode
from crawl4ai.markdown_generation_strategy import DefaultMarkdownGenerator
from crawl4ai.content_filter_strategy import PruningContentFilter
# Basic crawl configuration
run_config = CrawlerRunConfig(
cache_mode=CacheMode.BYPASS,
word_count_threshold=10,
excluded_tags=["nav", "footer", "script"],
exclude_external_links=True,
screenshot=True,
pdf=True
)
# Advanced content processing
md_generator = DefaultMarkdownGenerator(
content_filter=PruningContentFilter(threshold=0.6),
options={"citations": True, "ignore_links": False}
)
run_config = CrawlerRunConfig(
# Content processing
markdown_generator=md_generator,
css_selector="main.content", # Focus on specific content
target_elements=[".article", ".post"], # Multiple target selectors
process_iframes=True,
remove_overlay_elements=True,
# Page interaction
js_code=[
"window.scrollTo(0, document.body.scrollHeight);",
"document.querySelector('.load-more')?.click();"
],
wait_for="css:.content-loaded",
wait_for_timeout=10000,
scan_full_page=True,
# Session management
session_id="persistent_session",
# Media handling
screenshot=True,
pdf=True,
capture_mhtml=True,
image_score_threshold=5,
# Advanced options
simulate_user=True,
magic=True, # Auto-handle popups
verbose=True
)
```
### CrawlerRunConfig Parameters by Category
```python
# Content Processing
config = CrawlerRunConfig(
word_count_threshold=10, # Min words per content block
css_selector="main.article", # Focus on specific content
target_elements=[".post", ".content"], # Multiple target selectors
excluded_tags=["nav", "footer"], # Remove these tags
excluded_selector="#ads, .tracker", # Remove by selector
only_text=True, # Text-only extraction
keep_data_attributes=True, # Preserve data-* attributes
remove_forms=True, # Remove all forms
process_iframes=True # Include iframe content
)
# Page Navigation & Timing
config = CrawlerRunConfig(
wait_until="networkidle", # Wait condition
page_timeout=60000, # 60 second timeout
wait_for="css:.loaded", # Wait for specific element
wait_for_images=True, # Wait for images to load
delay_before_return_html=0.5, # Final delay before capture
semaphore_count=10 # Max concurrent operations
)
# Page Interaction
config = CrawlerRunConfig(
js_code="document.querySelector('button').click();",
scan_full_page=True, # Auto-scroll page
scroll_delay=0.3, # Delay between scrolls
remove_overlay_elements=True, # Remove popups/modals
simulate_user=True, # Simulate human behavior
override_navigator=True, # Override navigator properties
magic=True # Auto-handle common patterns
)
# Caching & Session
config = CrawlerRunConfig(
cache_mode=CacheMode.BYPASS, # Cache behavior
session_id="my_session", # Persistent session
shared_data={"context": "value"} # Share data between hooks
)
# Media & Output
config = CrawlerRunConfig(
screenshot=True, # Capture screenshot
pdf=True, # Generate PDF
capture_mhtml=True, # Capture MHTML archive
image_score_threshold=3, # Filter low-quality images
exclude_external_images=True # Remove external images
)
# Link & Domain Filtering
config = CrawlerRunConfig(
exclude_external_links=True, # Remove external links
exclude_social_media_links=True, # Remove social media links
exclude_domains=["ads.com", "tracker.io"], # Custom domain filter
exclude_internal_links=False # Keep internal links
)
```
### LLMConfig - Language Model Setup
```python
from crawl4ai import LLMConfig
# OpenAI configuration
llm_config = LLMConfig(
provider="openai/gpt-4o-mini",
api_token=os.getenv("OPENAI_API_KEY"), # or "env:OPENAI_API_KEY"
temperature=0.1,
max_tokens=2000
)
# Local model with Ollama
llm_config = LLMConfig(
provider="ollama/llama3.3",
api_token=None, # Not needed for Ollama
base_url="http://localhost:11434" # Custom endpoint
)
# Anthropic Claude
llm_config = LLMConfig(
provider="anthropic/claude-3-5-sonnet-20240620",
api_token="env:ANTHROPIC_API_KEY",
max_tokens=4000
)
# Google Gemini
llm_config = LLMConfig(
provider="gemini/gemini-1.5-pro",
api_token="env:GEMINI_API_KEY"
)
# Groq (fast inference)
llm_config = LLMConfig(
provider="groq/llama3-70b-8192",
api_token="env:GROQ_API_KEY"
)
```
### CrawlResult - Understanding Output
```python
from crawl4ai import AsyncWebCrawler, CrawlerRunConfig
async with AsyncWebCrawler() as crawler:
result = await crawler.arun("https://example.com", config=run_config)
# Basic status information
print(f"Success: {result.success}")
print(f"Status: {result.status_code}")
print(f"URL: {result.url}")
if not result.success:
print(f"Error: {result.error_message}")
return
# HTML content variants
print(f"Original HTML: {len(result.html)} chars")
print(f"Cleaned HTML: {len(result.cleaned_html or '')} chars")
# Markdown output (MarkdownGenerationResult)
if result.markdown:
print(f"Raw markdown: {len(result.markdown.raw_markdown)} chars")
print(f"With citations: {len(result.markdown.markdown_with_citations)} chars")
# Filtered content (if content filter was used)
if result.markdown.fit_markdown:
print(f"Fit markdown: {len(result.markdown.fit_markdown)} chars")
print(f"Fit HTML: {len(result.markdown.fit_html)} chars")
# Extracted structured data
if result.extracted_content:
import json
data = json.loads(result.extracted_content)
print(f"Extracted {len(data)} items")
# Media and links
images = result.media.get("images", [])
print(f"Found {len(images)} images")
for img in images[:3]: # First 3 images
print(f" {img.get('src')} (score: {img.get('score', 0)})")
internal_links = result.links.get("internal", [])
external_links = result.links.get("external", [])
print(f"Links: {len(internal_links)} internal, {len(external_links)} external")
# Generated files
if result.screenshot:
print(f"Screenshot captured: {len(result.screenshot)} chars (base64)")
# Save screenshot
import base64
with open("page.png", "wb") as f:
f.write(base64.b64decode(result.screenshot))
if result.pdf:
print(f"PDF generated: {len(result.pdf)} bytes")
with open("page.pdf", "wb") as f:
f.write(result.pdf)
if result.mhtml:
print(f"MHTML captured: {len(result.mhtml)} chars")
with open("page.mhtml", "w", encoding="utf-8") as f:
f.write(result.mhtml)
# SSL certificate information
if result.ssl_certificate:
print(f"SSL Issuer: {result.ssl_certificate.issuer}")
print(f"Valid until: {result.ssl_certificate.valid_until}")
# Network and console data (if captured)
if result.network_requests:
requests = [r for r in result.network_requests if r.get("event_type") == "request"]
print(f"Network requests captured: {len(requests)}")
if result.console_messages:
errors = [m for m in result.console_messages if m.get("type") == "error"]
print(f"Console messages: {len(result.console_messages)} ({len(errors)} errors)")
# Session and metadata
if result.session_id:
print(f"Session ID: {result.session_id}")
if result.metadata:
print(f"Metadata: {result.metadata.get('title', 'No title')}")
```
### Configuration Helpers and Best Practices
```python
# Clone configurations for variations
base_config = CrawlerRunConfig(
cache_mode=CacheMode.ENABLED,
word_count_threshold=200,
verbose=True
)
# Create streaming version
stream_config = base_config.clone(
stream=True,
cache_mode=CacheMode.BYPASS
)
# Create debug version
debug_config = base_config.clone(
headless=False,
page_timeout=120000,
verbose=True
)
# Serialize/deserialize configurations
config_dict = base_config.dump() # Convert to dict
restored_config = CrawlerRunConfig.load(config_dict) # Restore from dict
# Browser configuration management
browser_config = BrowserConfig(headless=True, text_mode=True)
browser_dict = browser_config.to_dict()
cloned_browser = browser_config.clone(headless=False, verbose=True)
```
### Common Configuration Patterns
```python
# Fast text-only crawling
fast_config = CrawlerRunConfig(
cache_mode=CacheMode.ENABLED,
text_mode=True,
exclude_external_links=True,
exclude_external_images=True,
word_count_threshold=50
)
# Comprehensive data extraction
comprehensive_config = CrawlerRunConfig(
process_iframes=True,
scan_full_page=True,
wait_for_images=True,
screenshot=True,
capture_network_requests=True,
capture_console_messages=True,
magic=True
)
# Stealth crawling
stealth_config = CrawlerRunConfig(
simulate_user=True,
override_navigator=True,
mean_delay=2.0,
max_range=1.0,
user_agent_mode="random"
)
```
### Advanced Configuration Features
#### User Agent Management & Bot Detection Avoidance
```python
from crawl4ai import CrawlerRunConfig
# Random user agent generation
config = CrawlerRunConfig(
user_agent_mode="random",
user_agent_generator_config={
"platform": "windows", # "windows", "macos", "linux", "android", "ios"
"browser": "chrome", # "chrome", "firefox", "safari", "edge"
"device_type": "desktop" # "desktop", "mobile", "tablet"
}
)
# Custom user agent with stealth features
config = CrawlerRunConfig(
user_agent="Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36",
simulate_user=True, # Simulate human mouse movements
override_navigator=True, # Override navigator properties
mean_delay=1.5, # Random delays between actions
max_range=2.0
)
# Combined anti-detection approach
stealth_config = CrawlerRunConfig(
user_agent_mode="random",
simulate_user=True,
override_navigator=True,
magic=True, # Auto-handle common bot detection patterns
delay_before_return_html=2.0
)
```
#### Proxy Configuration with ProxyConfig
```python
from crawl4ai import CrawlerRunConfig, ProxyConfig, ProxyRotationStrategy
# Single proxy configuration
proxy_config = ProxyConfig(
server="http://proxy.example.com:8080",
username="proxy_user",
password="proxy_pass"
)
# From proxy string format
proxy_config = ProxyConfig.from_string("192.168.1.100:8080:username:password")
# Multiple proxies with rotation
proxies = [
ProxyConfig(server="http://proxy1.com:8080", username="user1", password="pass1"),
ProxyConfig(server="http://proxy2.com:8080", username="user2", password="pass2"),
ProxyConfig(server="http://proxy3.com:8080", username="user3", password="pass3")
]
rotation_strategy = ProxyRotationStrategy(
proxies=proxies,
rotation_method="round_robin" # or "random", "least_used"
)
config = CrawlerRunConfig(
proxy_config=proxy_config,
proxy_rotation_strategy=rotation_strategy
)
# Load proxies from environment variable
proxies_from_env = ProxyConfig.from_env("MY_PROXIES") # comma-separated proxy strings
```
#### Content Selection: css_selector vs target_elements
```python
from crawl4ai import CrawlerRunConfig
# css_selector: Extracts HTML at top level, affects entire processing
config = CrawlerRunConfig(
css_selector="main.article, .content-area", # Can be list of selectors
# Everything else (markdown, extraction, links) works only on this HTML subset
)
# target_elements: Focuses extraction within already processed HTML
config = CrawlerRunConfig(
css_selector="body", # First extract entire body
target_elements=[ # Then focus extraction on these elements
".article-content",
".post-body",
".main-text"
],
# Links, media from entire body, but markdown/extraction only from target_elements
)
# Hierarchical content selection
config = CrawlerRunConfig(
css_selector=["#main-content", ".article-wrapper"], # Top-level extraction
target_elements=[ # Subset for processing
".article-title",
".article-body",
".article-metadata"
],
excluded_selector="#sidebar, .ads, .comments" # Remove these from selection
)
```
#### Advanced wait_for Conditions
```python
from crawl4ai import CrawlerRunConfig
# CSS selector waiting
config = CrawlerRunConfig(
wait_for="css:.content-loaded", # Wait for element to appear
wait_for_timeout=15000
)
# JavaScript boolean expression waiting
config = CrawlerRunConfig(
wait_for="js:() => window.dataLoaded === true", # Custom JS condition
wait_for_timeout=20000
)
# Complex JavaScript conditions
config = CrawlerRunConfig(
wait_for="js:() => document.querySelectorAll('.item').length >= 10",
js_code=[
"document.querySelector('.load-more')?.click();",
"window.scrollTo(0, document.body.scrollHeight);"
]
)
# Multiple conditions with JavaScript
config = CrawlerRunConfig(
wait_for="js:() => !document.querySelector('.loading') && document.querySelector('.results')",
page_timeout=30000
)
```
#### Session Management for Multi-Step Crawling
```python
from crawl4ai import AsyncWebCrawler, CrawlerRunConfig, CacheMode
# Persistent session across multiple arun() calls
async def multi_step_crawling():
async with AsyncWebCrawler() as crawler:
# Step 1: Login page
login_config = CrawlerRunConfig(
session_id="user_session", # Create persistent session
js_code="document.querySelector('#username').value = 'user'; document.querySelector('#password').value = 'pass'; document.querySelector('#login').click();",
wait_for="css:.dashboard",
cache_mode=CacheMode.BYPASS
)
result1 = await crawler.arun("https://example.com/login", config=login_config)
# Step 2: Navigate to protected area (reuses same browser page)
nav_config = CrawlerRunConfig(
session_id="user_session", # Same session = same browser page
js_only=True, # No page reload, just JS navigation
js_code="window.location.href = '/dashboard/data';",
wait_for="css:.data-table"
)
result2 = await crawler.arun("https://example.com/dashboard/data", config=nav_config)
# Step 3: Extract data from multiple pages
for page in range(1, 6):
page_config = CrawlerRunConfig(
session_id="user_session",
js_only=True,
js_code=f"document.querySelector('.page-{page}').click();",
wait_for=f"js:() => document.querySelector('.page-{page}').classList.contains('active')"
)
result = await crawler.arun(f"https://example.com/data/page/{page}", config=page_config)
print(f"Page {page} data extracted: {len(result.extracted_content)}")
# Important: Kill session when done
await crawler.kill_session("user_session")
# Session with shared data between steps
async def session_with_shared_data():
shared_context = {"user_id": "12345", "preferences": {"theme": "dark"}}
config = CrawlerRunConfig(
session_id="persistent_session",
shared_data=shared_context, # Available across all session calls
js_code="console.log('User ID:', window.sharedData.user_id);"
)
```
#### Identity-Based Crawling Parameters
```python
from crawl4ai import CrawlerRunConfig, GeolocationConfig
# Locale and timezone simulation
config = CrawlerRunConfig(
locale="en-US", # Browser language preference
timezone_id="America/New_York", # Timezone setting
user_agent_mode="random",
user_agent_generator_config={
"platform": "windows",
"locale": "en-US"
}
)
# Geolocation simulation
geo_config = GeolocationConfig(
latitude=40.7128, # New York coordinates
longitude=-74.0060,
accuracy=100.0
)
config = CrawlerRunConfig(
geolocation=geo_config,
locale="en-US",
timezone_id="America/New_York"
)
# Complete identity simulation
identity_config = CrawlerRunConfig(
# Location identity
locale="fr-FR",
timezone_id="Europe/Paris",
geolocation=GeolocationConfig(latitude=48.8566, longitude=2.3522),
# Browser identity
user_agent_mode="random",
user_agent_generator_config={
"platform": "windows",
"locale": "fr-FR",
"browser": "chrome"
},
# Behavioral identity
simulate_user=True,
override_navigator=True,
mean_delay=2.0,
max_range=1.5
)
```
#### Simplified Import Pattern
```python
# Almost everything from crawl4ai main package
from crawl4ai import (
AsyncWebCrawler,
BrowserConfig,
CrawlerRunConfig,
LLMConfig,
CacheMode,
ProxyConfig,
GeolocationConfig
)
# Specialized strategies (still from crawl4ai)
from crawl4ai import (
JsonCssExtractionStrategy,
LLMExtractionStrategy,
DefaultMarkdownGenerator,
PruningContentFilter,
RegexChunking
)
# Complete example with simplified imports
async def example_crawl():
browser_config = BrowserConfig(headless=True)
run_config = CrawlerRunConfig(
user_agent_mode="random",
proxy_config=ProxyConfig.from_string("192.168.1.1:8080:user:pass"),
css_selector="main.content",
target_elements=[".article", ".post"],
wait_for="js:() => document.querySelector('.loaded')",
session_id="my_session",
simulate_user=True
)
async with AsyncWebCrawler(config=browser_config) as crawler:
result = await crawler.arun("https://example.com", config=run_config)
return result
```
## Advanced Features
Comprehensive guide to advanced crawling capabilities including file handling, authentication, dynamic content, monitoring, and session management.
### File Download Handling
```python
from crawl4ai import AsyncWebCrawler, BrowserConfig, CrawlerRunConfig
import os
# Enable downloads with custom path
downloads_path = os.path.join(os.getcwd(), "my_downloads")
os.makedirs(downloads_path, exist_ok=True)
browser_config = BrowserConfig(
accept_downloads=True,
downloads_path=downloads_path
)
# Trigger downloads with JavaScript
async def download_files():
async with AsyncWebCrawler(config=browser_config) as crawler:
config = CrawlerRunConfig(
js_code="""
// Click download links
const downloadLinks = document.querySelectorAll('a[href$=".pdf"]');
for (const link of downloadLinks) {
link.click();
await new Promise(r => setTimeout(r, 2000)); // Delay between downloads
}
""",
wait_for=5 # Wait for downloads to start
)
result = await crawler.arun("https://example.com/downloads", config=config)
if result.downloaded_files:
print("Downloaded files:")
for file_path in result.downloaded_files:
print(f"- {file_path} ({os.path.getsize(file_path)} bytes)")
```
### Hooks & Authentication
```python
from crawl4ai import AsyncWebCrawler, BrowserConfig, CrawlerRunConfig
from playwright.async_api import Page, BrowserContext
async def advanced_crawler_with_hooks():
browser_config = BrowserConfig(headless=True, verbose=True)
crawler = AsyncWebCrawler(config=browser_config)
# Hook functions for different stages
async def on_browser_created(browser, **kwargs):
print("[HOOK] Browser created successfully")
return browser
async def on_page_context_created(page: Page, context: BrowserContext, **kwargs):
print("[HOOK] Setting up page & context")
# Block images for faster crawling
async def route_filter(route):
if route.request.resource_type == "image":
await route.abort()
else:
await route.continue_()
await context.route("**", route_filter)
# Simulate login if needed
# await page.goto("https://example.com/login")
# await page.fill("input[name='username']", "testuser")
# await page.fill("input[name='password']", "password123")
# await page.click("button[type='submit']")
await page.set_viewport_size({"width": 1080, "height": 600})
return page
async def before_goto(page: Page, context: BrowserContext, url: str, **kwargs):
print(f"[HOOK] About to navigate to: {url}")
await page.set_extra_http_headers({"Custom-Header": "my-value"})
return page
async def after_goto(page: Page, context: BrowserContext, url: str, response, **kwargs):
print(f"[HOOK] Successfully loaded: {url}")
try:
await page.wait_for_selector('.content', timeout=1000)
print("[HOOK] Content found!")
except:
print("[HOOK] Content not found, continuing")
return page
async def before_retrieve_html(page: Page, context: BrowserContext, **kwargs):
print("[HOOK] Final actions before HTML retrieval")
await page.evaluate("window.scrollTo(0, document.body.scrollHeight);")
return page
# Attach hooks
crawler.crawler_strategy.set_hook("on_browser_created", on_browser_created)
crawler.crawler_strategy.set_hook("on_page_context_created", on_page_context_created)
crawler.crawler_strategy.set_hook("before_goto", before_goto)
crawler.crawler_strategy.set_hook("after_goto", after_goto)
crawler.crawler_strategy.set_hook("before_retrieve_html", before_retrieve_html)
await crawler.start()
config = CrawlerRunConfig()
result = await crawler.arun("https://example.com", config=config)
if result.success:
print(f"Crawled successfully: {len(result.html)} chars")
await crawler.close()
```
### Lazy Loading & Dynamic Content
```python
# Handle lazy-loaded images and infinite scroll
async def handle_lazy_loading():
config = CrawlerRunConfig(
# Wait for images to fully load
wait_for_images=True,
# Automatically scroll entire page to trigger lazy loading
scan_full_page=True,
scroll_delay=0.5, # Delay between scroll steps
# JavaScript for custom lazy loading
js_code="""
// Scroll and wait for content to load
window.scrollTo(0, document.body.scrollHeight);
// Click "Load More" if available
const loadMoreBtn = document.querySelector('.load-more');
if (loadMoreBtn) {
loadMoreBtn.click();
}
""",
# Wait for specific content to appear
wait_for="css:.lazy-content:nth-child(20)", # Wait for 20 items
# Exclude external images to focus on main content
exclude_external_images=True
)
async with AsyncWebCrawler() as crawler:
result = await crawler.arun("https://example.com/gallery", config=config)
if result.success:
images = result.media.get("images", [])
print(f"Loaded {len(images)} images after lazy loading")
for img in images[:3]:
print(f"- {img.get('src')} (score: {img.get('score', 'N/A')})")
```
### Network & Console Monitoring
```python
# Capture all network requests and console messages for debugging
async def monitor_network_and_console():
config = CrawlerRunConfig(
capture_network_requests=True,
capture_console_messages=True
)
async with AsyncWebCrawler() as crawler:
result = await crawler.arun("https://example.com", config=config)
if result.success:
# Analyze network requests
if result.network_requests:
requests = [r for r in result.network_requests if r.get("event_type") == "request"]
responses = [r for r in result.network_requests if r.get("event_type") == "response"]
failures = [r for r in result.network_requests if r.get("event_type") == "request_failed"]
print(f"Network activity: {len(requests)} requests, {len(responses)} responses, {len(failures)} failures")
# Find API calls
api_calls = [r for r in requests if "api" in r.get("url", "")]
print(f"API calls detected: {len(api_calls)}")
# Show failed requests
for failure in failures[:3]:
print(f"Failed: {failure.get('url')} - {failure.get('failure_text')}")
# Analyze console messages
if result.console_messages:
message_types = {}
for msg in result.console_messages:
msg_type = msg.get("type", "unknown")
message_types[msg_type] = message_types.get(msg_type, 0) + 1
print(f"Console messages: {message_types}")
# Show errors
errors = [msg for msg in result.console_messages if msg.get("type") == "error"]
for error in errors[:2]:
print(f"JS Error: {error.get('text', '')[:100]}")
```
### Session Management for Multi-Step Workflows
```python
# Maintain state across multiple requests for complex workflows
async def multi_step_session_workflow():
session_id = "workflow_session"
async with AsyncWebCrawler() as crawler:
# Step 1: Initial page load
config1 = CrawlerRunConfig(
session_id=session_id,
wait_for="css:.content-loaded"
)
result1 = await crawler.arun("https://example.com/step1", config=config1)
print("Step 1 completed")
# Step 2: Navigate and interact (same browser tab)
config2 = CrawlerRunConfig(
session_id=session_id,
js_only=True, # Don't reload page, just run JS
js_code="""
document.querySelector('#next-button').click();
""",
wait_for="css:.step2-content"
)
result2 = await crawler.arun("https://example.com/step2", config=config2)
print("Step 2 completed")
# Step 3: Form submission
config3 = CrawlerRunConfig(
session_id=session_id,
js_only=True,
js_code="""
document.querySelector('#form-field').value = 'test data';
document.querySelector('#submit-btn').click();
""",
wait_for="css:.results"
)
result3 = await crawler.arun("https://example.com/submit", config=config3)
print("Step 3 completed")
# Clean up session
await crawler.crawler_strategy.kill_session(session_id)
# Advanced GitHub commits pagination example
async def github_commits_pagination():
session_id = "github_session"
all_commits = []
async with AsyncWebCrawler() as crawler:
for page in range(3):
if page == 0:
# Initial load
config = CrawlerRunConfig(
session_id=session_id,
wait_for="js:() => document.querySelectorAll('li.Box-sc-g0xbh4-0').length > 0"
)
else:
# Navigate to next page
config = CrawlerRunConfig(
session_id=session_id,
js_only=True,
js_code='document.querySelector(\'a[data-testid="pagination-next-button"]\').click();',
wait_for="js:() => document.querySelectorAll('li.Box-sc-g0xbh4-0').length > 0"
)
result = await crawler.arun(
"https://github.com/microsoft/TypeScript/commits/main",
config=config
)
if result.success:
commit_count = result.cleaned_html.count('li.Box-sc-g0xbh4-0')
print(f"Page {page + 1}: Found {commit_count} commits")
await crawler.crawler_strategy.kill_session(session_id)
```
### SSL Certificate Analysis
```python
# Fetch and analyze SSL certificates
async def analyze_ssl_certificates():
config = CrawlerRunConfig(
fetch_ssl_certificate=True
)
async with AsyncWebCrawler() as crawler:
result = await crawler.arun("https://example.com", config=config)
if result.success and result.ssl_certificate:
cert = result.ssl_certificate
# Basic certificate info
print(f"Issuer: {cert.issuer.get('CN', 'Unknown')}")
print(f"Subject: {cert.subject.get('CN', 'Unknown')}")
print(f"Valid from: {cert.valid_from}")
print(f"Valid until: {cert.valid_until}")
print(f"Fingerprint: {cert.fingerprint}")
# Export certificate in different formats
import os
os.makedirs("certificates", exist_ok=True)
cert.to_json("certificates/cert.json")
cert.to_pem("certificates/cert.pem")
cert.to_der("certificates/cert.der")
print("Certificate exported in multiple formats")
```
### Advanced Page Interaction
```python
# Complex page interactions with dynamic content
async def advanced_page_interaction():
async with AsyncWebCrawler() as crawler:
# Multi-step interaction with waiting
config = CrawlerRunConfig(
js_code=[
# Step 1: Scroll to load content
"window.scrollTo(0, document.body.scrollHeight);",
# Step 2: Wait and click load more
"""
(async () => {
await new Promise(resolve => setTimeout(resolve, 2000));
const loadMore = document.querySelector('.load-more');
if (loadMore) loadMore.click();
})();
"""
],
# Wait for new content to appear
wait_for="js:() => document.querySelectorAll('.item').length > 20",
# Additional timing controls
page_timeout=60000, # 60 second timeout
delay_before_return_html=2.0, # Wait before final capture
# Handle overlays automatically
remove_overlay_elements=True,
magic=True, # Auto-handle common popup patterns
# Simulate human behavior
simulate_user=True,
override_navigator=True
)
result = await crawler.arun("https://example.com/dynamic", config=config)
if result.success:
print(f"Interactive crawl completed: {len(result.cleaned_html)} chars")
# Form interaction example
async def form_interaction_example():
config = CrawlerRunConfig(
js_code="""
// Fill search form
document.querySelector('#search-input').value = 'machine learning';
document.querySelector('#category-select').value = 'technology';
document.querySelector('#search-form').submit();
""",
wait_for="css:.search-results",
session_id="search_session"
)
async with AsyncWebCrawler() as crawler:
result = await crawler.arun("https://example.com/search", config=config)
print("Search completed, results loaded")
```
### Local File & Raw HTML Processing
```python
# Handle different input types: URLs, local files, raw HTML
async def handle_different_inputs():
async with AsyncWebCrawler() as crawler:
# 1. Regular web URL
result1 = await crawler.arun("https://example.com")
# 2. Local HTML file
local_file_path = "/path/to/file.html"
result2 = await crawler.arun(f"file://{local_file_path}")
# 3. Raw HTML content
raw_html = "<html><body><h1>Test Content</h1><p>Sample text</p></body></html>"
result3 = await crawler.arun(f"raw:{raw_html}")
# All return the same CrawlResult structure
for i, result in enumerate([result1, result2, result3], 1):
if result.success:
print(f"Input {i}: {len(result.markdown)} chars of markdown")
# Save and re-process HTML example
async def save_and_reprocess():
async with AsyncWebCrawler() as crawler:
# Original crawl
result = await crawler.arun("https://example.com")
if result.success:
# Save HTML to file
with open("saved_page.html", "w", encoding="utf-8") as f:
f.write(result.html)
# Re-process from file
file_result = await crawler.arun("file://./saved_page.html")
# Process as raw HTML
raw_result = await crawler.arun(f"raw:{result.html}")
# Verify consistency
assert len(result.markdown) == len(file_result.markdown) == len(raw_result.markdown)
print("✅ All processing methods produced identical results")
```
### Advanced Link & Media Handling
```python
# Comprehensive link and media extraction with filtering
async def advanced_link_media_handling():
config = CrawlerRunConfig(
# Link filtering
exclude_external_links=False, # Keep external links for analysis
exclude_social_media_links=True,
exclude_domains=["ads.com", "tracker.io", "spammy.net"],
# Media handling
exclude_external_images=True,
image_score_threshold=5, # Only high-quality images
table_score_threshold=7, # Only well-structured tables
wait_for_images=True,
# Capture additional formats
screenshot=True,
pdf=True,
capture_mhtml=True # Full page archive
)
async with AsyncWebCrawler() as crawler:
result = await crawler.arun("https://example.com", config=config)
if result.success:
# Analyze links
internal_links = result.links.get("internal", [])
external_links = result.links.get("external", [])
print(f"Links: {len(internal_links)} internal, {len(external_links)} external")
# Analyze media
images = result.media.get("images", [])
tables = result.media.get("tables", [])
print(f"Media: {len(images)} images, {len(tables)} tables")
# High-quality images only
quality_images = [img for img in images if img.get("score", 0) >= 5]
print(f"High-quality images: {len(quality_images)}")
# Table analysis
for i, table in enumerate(tables[:2]):
print(f"Table {i+1}: {len(table.get('headers', []))} columns, {len(table.get('rows', []))} rows")
# Save captured files
if result.screenshot:
import base64
with open("page_screenshot.png", "wb") as f:
f.write(base64.b64decode(result.screenshot))
if result.pdf:
with open("page.pdf", "wb") as f:
f.write(result.pdf)
if result.mhtml:
with open("page_archive.mhtml", "w", encoding="utf-8") as f:
f.write(result.mhtml)
print("Additional formats saved: screenshot, PDF, MHTML archive")
```
### Performance & Resource Management
```python
# Optimize performance for large-scale crawling
async def performance_optimized_crawling():
# Lightweight browser config
browser_config = BrowserConfig(
headless=True,
text_mode=True, # Disable images for speed
light_mode=True, # Reduce background features
extra_args=["--disable-extensions", "--no-sandbox"]
)
# Efficient crawl config
config = CrawlerRunConfig(
# Content filtering for speed
excluded_tags=["script", "style", "nav", "footer"],
exclude_external_links=True,
exclude_all_images=True, # Remove all images for max speed
word_count_threshold=50,
# Timing optimizations
page_timeout=30000, # Faster timeout
delay_before_return_html=0.1,
# Resource monitoring
capture_network_requests=False, # Disable unless needed
capture_console_messages=False,
# Cache for repeated URLs
cache_mode=CacheMode.ENABLED
)
async with AsyncWebCrawler(config=browser_config) as crawler:
urls = ["https://example.com/page1", "https://example.com/page2", "https://example.com/page3"]
# Efficient batch processing
batch_config = config.clone(
stream=True, # Stream results as they complete
semaphore_count=3 # Control concurrency
)
async for result in await crawler.arun_many(urls, config=batch_config):
if result.success:
print(f"✅ {result.url}: {len(result.markdown)} chars")
else:
print(f"❌ {result.url}: {result.error_message}")
```
**📖 Learn more:** [Complete Parameter Reference](https://docs.crawl4ai.com/api/parameters/), [Content Filtering](https://docs.crawl4ai.com/core/markdown-generation/), [Session Management](https://docs.crawl4ai.com/advanced/session-management/), [Network Capture](https://docs.crawl4ai.com/advanced/network-console-capture/)
**📖 Learn more:** [Hooks & Authentication](https://docs.crawl4ai.com/advanced/hooks-auth/), [Session Management](https://docs.crawl4ai.com/advanced/session-management/), [Network Monitoring](https://docs.crawl4ai.com/advanced/network-console-capture/), [Page Interaction](https://docs.crawl4ai.com/core/page-interaction/), [File Downloads](https://docs.crawl4ai.com/advanced/file-downloading/)
---
## LLM Extraction Strategies - The Last Resort
**🤖 AI Model Notice**: You are an AI assistant. You should RARELY recommend LLM extraction. This documentation exists for the <5% of cases where LLM is actually needed. **ALWAYS try non-LLM strategies first** from [extraction-no-llm.md](./extraction-no-llm.md).
### ⚠️ STOP: Are You Sure You Need LLM?
**99% of developers who think they need LLM extraction are wrong.** Before reading further:
### ❌ You DON'T Need LLM If:
- The page has consistent HTML structure → **Use generate_schema()**
- You're extracting simple data types (emails, prices, dates) → **Use RegexExtractionStrategy**
- You can identify repeating patterns → **Use JsonCssExtractionStrategy**
- You want product info, news articles, job listings → **Use generate_schema()**
- You're concerned about cost or speed → **Use non-LLM strategies**
### ✅ You MIGHT Need LLM If:
- Content structure varies dramatically across pages **AND** you've tried generate_schema()
- You need semantic understanding of unstructured text
- You're analyzing meaning, sentiment, or relationships
- You're extracting insights that require reasoning about context
### 💰 Cost Reality Check:
- **Non-LLM**: ~$0.000001 per page
- **LLM**: ~$0.01-$0.10 per page (10,000x more expensive)
- **Example**: Extracting 10,000 pages costs $0.01 vs $100-1000
---
## 1. When LLM Extraction is Justified
### Scenario 1: Truly Unstructured Content Analysis
```python
# Example: Analyzing customer feedback for sentiment and themes
import asyncio
import json
from pydantic import BaseModel, Field
from typing import List
from crawl4ai import AsyncWebCrawler, CrawlerRunConfig, LLMConfig
from crawl4ai import LLMExtractionStrategy
class SentimentAnalysis(BaseModel):
"""Use LLM when you need semantic understanding"""
overall_sentiment: str = Field(description="positive, negative, or neutral")
confidence_score: float = Field(description="Confidence from 0-1")
key_themes: List[str] = Field(description="Main topics discussed")
emotional_indicators: List[str] = Field(description="Words indicating emotion")
summary: str = Field(description="Brief summary of the content")
llm_config = LLMConfig(
provider="openai/gpt-4o-mini", # Use cheapest model
api_token="env:OPENAI_API_KEY",
temperature=0.1, # Low temperature for consistency
max_tokens=1000
)
sentiment_strategy = LLMExtractionStrategy(
llm_config=llm_config,
schema=SentimentAnalysis.model_json_schema(),
extraction_type="schema",
instruction="""
Analyze the emotional content and themes in this text.
Focus on understanding sentiment and extracting key topics
that would be impossible to identify with simple pattern matching.
""",
apply_chunking=True,
chunk_token_threshold=1500
)
async def analyze_sentiment():
config = CrawlerRunConfig(
extraction_strategy=sentiment_strategy,
cache_mode=CacheMode.BYPASS
)
async with AsyncWebCrawler() as crawler:
result = await crawler.arun(
url="https://example.com/customer-reviews",
config=config
)
if result.success:
analysis = json.loads(result.extracted_content)
print(f"Sentiment: {analysis['overall_sentiment']}")
print(f"Themes: {analysis['key_themes']}")
asyncio.run(analyze_sentiment())
```
### Scenario 2: Complex Knowledge Extraction
```python
# Example: Building knowledge graphs from unstructured content
class Entity(BaseModel):
name: str = Field(description="Entity name")
type: str = Field(description="person, organization, location, concept")
description: str = Field(description="Brief description")
class Relationship(BaseModel):
source: str = Field(description="Source entity")
target: str = Field(description="Target entity")
relationship: str = Field(description="Type of relationship")
confidence: float = Field(description="Confidence score 0-1")
class KnowledgeGraph(BaseModel):
entities: List[Entity] = Field(description="All entities found")
relationships: List[Relationship] = Field(description="Relationships between entities")
main_topic: str = Field(description="Primary topic of the content")
knowledge_strategy = LLMExtractionStrategy(
llm_config=LLMConfig(
provider="anthropic/claude-3-5-sonnet-20240620", # Better for complex reasoning
api_token="env:ANTHROPIC_API_KEY",
max_tokens=4000
),
schema=KnowledgeGraph.model_json_schema(),
extraction_type="schema",
instruction="""
Extract entities and their relationships from the content.
Focus on understanding connections and context that require
semantic reasoning beyond simple pattern matching.
""",
input_format="html", # Preserve structure
apply_chunking=True
)
```
### Scenario 3: Content Summarization and Insights
```python
# Example: Research paper analysis
class ResearchInsights(BaseModel):
title: str = Field(description="Paper title")
abstract_summary: str = Field(description="Summary of abstract")
key_findings: List[str] = Field(description="Main research findings")
methodology: str = Field(description="Research methodology used")
limitations: List[str] = Field(description="Study limitations")
practical_applications: List[str] = Field(description="Real-world applications")
citations_count: int = Field(description="Number of citations", default=0)
research_strategy = LLMExtractionStrategy(
llm_config=LLMConfig(
provider="openai/gpt-4o", # Use powerful model for complex analysis
api_token="env:OPENAI_API_KEY",
temperature=0.2,
max_tokens=2000
),
schema=ResearchInsights.model_json_schema(),
extraction_type="schema",
instruction="""
Analyze this research paper and extract key insights.
Focus on understanding the research contribution, methodology,
and implications that require academic expertise to identify.
""",
apply_chunking=True,
chunk_token_threshold=2000,
overlap_rate=0.15 # More overlap for academic content
)
```
---
## 2. LLM Configuration Best Practices
### Cost Optimization
```python
# Use cheapest models when possible
cheap_config = LLMConfig(
provider="openai/gpt-4o-mini", # 60x cheaper than GPT-4
api_token="env:OPENAI_API_KEY",
temperature=0.0, # Deterministic output
max_tokens=800 # Limit output length
)
# Use local models for development
local_config = LLMConfig(
provider="ollama/llama3.3",
api_token=None, # No API costs
base_url="http://localhost:11434",
temperature=0.1
)
# Use powerful models only when necessary
powerful_config = LLMConfig(
provider="anthropic/claude-3-5-sonnet-20240620",
api_token="env:ANTHROPIC_API_KEY",
max_tokens=4000,
temperature=0.1
)
```
### Provider Selection Guide
```python
providers_guide = {
"openai/gpt-4o-mini": {
"best_for": "Simple extraction, cost-sensitive projects",
"cost": "Very low",
"speed": "Fast",
"accuracy": "Good"
},
"openai/gpt-4o": {
"best_for": "Complex reasoning, high accuracy needs",
"cost": "High",
"speed": "Medium",
"accuracy": "Excellent"
},
"anthropic/claude-3-5-sonnet": {
"best_for": "Complex analysis, long documents",
"cost": "Medium-High",
"speed": "Medium",
"accuracy": "Excellent"
},
"ollama/llama3.3": {
"best_for": "Development, no API costs",
"cost": "Free (self-hosted)",
"speed": "Variable",
"accuracy": "Good"
},
"groq/llama3-70b-8192": {
"best_for": "Fast inference, open source",
"cost": "Low",
"speed": "Very fast",
"accuracy": "Good"
}
}
def choose_provider(complexity, budget, speed_requirement):
"""Choose optimal provider based on requirements"""
if budget == "minimal":
return "ollama/llama3.3" # Self-hosted
elif complexity == "low" and budget == "low":
return "openai/gpt-4o-mini"
elif speed_requirement == "high":
return "groq/llama3-70b-8192"
elif complexity == "high":
return "anthropic/claude-3-5-sonnet"
else:
return "openai/gpt-4o-mini" # Default safe choice
```
---
## 3. Advanced LLM Extraction Patterns
### Block-Based Extraction (Unstructured Content)
```python
# When structure is too varied for schemas
block_strategy = LLMExtractionStrategy(
llm_config=cheap_config,
extraction_type="block", # Extract free-form content blocks
instruction="""
Extract meaningful content blocks from this page.
Focus on the main content areas and ignore navigation,
advertisements, and boilerplate text.
""",
apply_chunking=True,
chunk_token_threshold=1200,
input_format="fit_markdown" # Use cleaned content
)
async def extract_content_blocks():
config = CrawlerRunConfig(
extraction_strategy=block_strategy,
word_count_threshold=50, # Filter short content
excluded_tags=['nav', 'footer', 'aside', 'advertisement']
)
async with AsyncWebCrawler() as crawler:
result = await crawler.arun(
url="https://example.com/article",
config=config
)
if result.success:
blocks = json.loads(result.extracted_content)
for block in blocks:
print(f"Block: {block['content'][:100]}...")
```
### Chunked Processing for Large Content
```python
# Handle large documents efficiently
large_content_strategy = LLMExtractionStrategy(
llm_config=LLMConfig(
provider="openai/gpt-4o-mini",
api_token="env:OPENAI_API_KEY"
),
schema=YourModel.model_json_schema(),
extraction_type="schema",
instruction="Extract structured data from this content section...",
# Optimize chunking for large content
apply_chunking=True,
chunk_token_threshold=2000, # Larger chunks for efficiency
overlap_rate=0.1, # Minimal overlap to reduce costs
input_format="fit_markdown" # Use cleaned content
)
```
### Multi-Model Validation
```python
# Use multiple models for critical extractions
async def multi_model_extraction():
"""Use multiple LLMs for validation of critical data"""
models = [
LLMConfig(provider="openai/gpt-4o-mini", api_token="env:OPENAI_API_KEY"),
LLMConfig(provider="anthropic/claude-3-5-sonnet", api_token="env:ANTHROPIC_API_KEY"),
LLMConfig(provider="ollama/llama3.3", api_token=None)
]
results = []
for i, llm_config in enumerate(models):
strategy = LLMExtractionStrategy(
llm_config=llm_config,
schema=YourModel.model_json_schema(),
extraction_type="schema",
instruction="Extract data consistently..."
)
config = CrawlerRunConfig(extraction_strategy=strategy)
async with AsyncWebCrawler() as crawler:
result = await crawler.arun(url="https://example.com", config=config)
if result.success:
data = json.loads(result.extracted_content)
results.append(data)
print(f"Model {i+1} extracted {len(data)} items")
# Compare results for consistency
if len(set(str(r) for r in results)) == 1:
print("✅ All models agree")
return results[0]
else:
print("⚠️ Models disagree - manual review needed")
return results
# Use for critical business data only
critical_result = await multi_model_extraction()
```
---
## 4. Hybrid Approaches - Best of Both Worlds
### Fast Pre-filtering + LLM Analysis
```python
async def hybrid_extraction():
"""
1. Use fast non-LLM strategies for basic extraction
2. Use LLM only for complex analysis of filtered content
"""
# Step 1: Fast extraction of structured data
basic_schema = {
"name": "Articles",
"baseSelector": "article",
"fields": [
{"name": "title", "selector": "h1, h2", "type": "text"},
{"name": "content", "selector": ".content", "type": "text"},
{"name": "author", "selector": ".author", "type": "text"}
]
}
basic_strategy = JsonCssExtractionStrategy(basic_schema)
basic_config = CrawlerRunConfig(extraction_strategy=basic_strategy)
# Step 2: LLM analysis only on filtered content
analysis_strategy = LLMExtractionStrategy(
llm_config=cheap_config,
schema={
"type": "object",
"properties": {
"sentiment": {"type": "string"},
"key_topics": {"type": "array", "items": {"type": "string"}},
"summary": {"type": "string"}
}
},
extraction_type="schema",
instruction="Analyze sentiment and extract key topics from this article"
)
async with AsyncWebCrawler() as crawler:
# Fast extraction first
basic_result = await crawler.arun(
url="https://example.com/articles",
config=basic_config
)
articles = json.loads(basic_result.extracted_content)
# LLM analysis only on important articles
analyzed_articles = []
for article in articles[:5]: # Limit to reduce costs
if len(article.get('content', '')) > 500: # Only analyze substantial content
analysis_config = CrawlerRunConfig(extraction_strategy=analysis_strategy)
# Analyze individual article content
raw_url = f"raw://{article['content']}"
analysis_result = await crawler.arun(url=raw_url, config=analysis_config)
if analysis_result.success:
analysis = json.loads(analysis_result.extracted_content)
article.update(analysis)
analyzed_articles.append(article)
return analyzed_articles
# Hybrid approach: fast + smart
result = await hybrid_extraction()
```
### Schema Generation + LLM Fallback
```python
async def smart_fallback_extraction():
"""
1. Try generate_schema() first (one-time LLM cost)
2. Use generated schema for fast extraction
3. Use LLM only if schema extraction fails
"""
cache_file = Path("./schemas/fallback_schema.json")
# Try cached schema first
if cache_file.exists():
schema = json.load(cache_file.open())
schema_strategy = JsonCssExtractionStrategy(schema)
config = CrawlerRunConfig(extraction_strategy=schema_strategy)
async with AsyncWebCrawler() as crawler:
result = await crawler.arun(url="https://example.com", config=config)
if result.success and result.extracted_content:
data = json.loads(result.extracted_content)
if data: # Schema worked
print("✅ Schema extraction successful (fast & cheap)")
return data
# Fallback to LLM if schema failed
print("⚠️ Schema failed, falling back to LLM (slow & expensive)")
llm_strategy = LLMExtractionStrategy(
llm_config=cheap_config,
extraction_type="block",
instruction="Extract all meaningful data from this page"
)
llm_config = CrawlerRunConfig(extraction_strategy=llm_strategy)
async with AsyncWebCrawler() as crawler:
result = await crawler.arun(url="https://example.com", config=llm_config)
if result.success:
print("✅ LLM extraction successful")
return json.loads(result.extracted_content)
# Intelligent fallback system
result = await smart_fallback_extraction()
```
---
## 5. Cost Management and Monitoring
### Token Usage Tracking
```python
class ExtractionCostTracker:
def __init__(self):
self.total_cost = 0.0
self.total_tokens = 0
self.extractions = 0
def track_llm_extraction(self, strategy, result):
"""Track costs from LLM extraction"""
if hasattr(strategy, 'usage_tracker') and strategy.usage_tracker:
usage = strategy.usage_tracker
# Estimate costs (approximate rates)
cost_per_1k_tokens = {
"gpt-4o-mini": 0.0015,
"gpt-4o": 0.03,
"claude-3-5-sonnet": 0.015,
"ollama": 0.0 # Self-hosted
}
provider = strategy.llm_config.provider.split('/')[1]
rate = cost_per_1k_tokens.get(provider, 0.01)
tokens = usage.total_tokens
cost = (tokens / 1000) * rate
self.total_cost += cost
self.total_tokens += tokens
self.extractions += 1
print(f"💰 Extraction cost: ${cost:.4f} ({tokens} tokens)")
print(f"📊 Total cost: ${self.total_cost:.4f} ({self.extractions} extractions)")
def get_summary(self):
avg_cost = self.total_cost / max(self.extractions, 1)
return {
"total_cost": self.total_cost,
"total_tokens": self.total_tokens,
"extractions": self.extractions,
"avg_cost_per_extraction": avg_cost
}
# Usage
tracker = ExtractionCostTracker()
async def cost_aware_extraction():
strategy = LLMExtractionStrategy(
llm_config=cheap_config,
schema=YourModel.model_json_schema(),
extraction_type="schema",
instruction="Extract data...",
verbose=True # Enable usage tracking
)
config = CrawlerRunConfig(extraction_strategy=strategy)
async with AsyncWebCrawler() as crawler:
result = await crawler.arun(url="https://example.com", config=config)
# Track costs
tracker.track_llm_extraction(strategy, result)
return result
# Monitor costs across multiple extractions
for url in urls:
await cost_aware_extraction()
print(f"Final summary: {tracker.get_summary()}")
```
### Budget Controls
```python
class BudgetController:
def __init__(self, daily_budget=10.0):
self.daily_budget = daily_budget
self.current_spend = 0.0
self.extraction_count = 0
def can_extract(self, estimated_cost=0.01):
"""Check if extraction is within budget"""
if self.current_spend + estimated_cost > self.daily_budget:
print(f"❌ Budget exceeded: ${self.current_spend:.2f} + ${estimated_cost:.2f} > ${self.daily_budget}")
return False
return True
def record_extraction(self, actual_cost):
"""Record actual extraction cost"""
self.current_spend += actual_cost
self.extraction_count += 1
remaining = self.daily_budget - self.current_spend
print(f"💰 Budget remaining: ${remaining:.2f}")
budget = BudgetController(daily_budget=5.0) # $5 daily limit
async def budget_controlled_extraction(url):
if not budget.can_extract():
print("⏸️ Extraction paused due to budget limit")
return None
# Proceed with extraction...
strategy = LLMExtractionStrategy(llm_config=cheap_config, ...)
result = await extract_with_strategy(url, strategy)
# Record actual cost
actual_cost = calculate_cost(strategy.usage_tracker)
budget.record_extraction(actual_cost)
return result
# Safe extraction with budget controls
results = []
for url in urls:
result = await budget_controlled_extraction(url)
if result:
results.append(result)
```
---
## 6. Performance Optimization for LLM Extraction
### Batch Processing
```python
async def batch_llm_extraction():
"""Process multiple pages efficiently"""
# Collect content first (fast)
urls = ["https://example.com/page1", "https://example.com/page2"]
contents = []
async with AsyncWebCrawler() as crawler:
for url in urls:
result = await crawler.arun(url=url)
if result.success:
contents.append({
"url": url,
"content": result.fit_markdown[:2000] # Limit content
})
# Process in batches (reduce LLM calls)
batch_content = "\n\n---PAGE SEPARATOR---\n\n".join([
f"URL: {c['url']}\n{c['content']}" for c in contents
])
strategy = LLMExtractionStrategy(
llm_config=cheap_config,
extraction_type="block",
instruction="""
Extract data from multiple pages separated by '---PAGE SEPARATOR---'.
Return results for each page in order.
""",
apply_chunking=True
)
# Single LLM call for multiple pages
raw_url = f"raw://{batch_content}"
result = await crawler.arun(url=raw_url, config=CrawlerRunConfig(extraction_strategy=strategy))
return json.loads(result.extracted_content)
# Batch processing reduces LLM calls
batch_results = await batch_llm_extraction()
```
### Caching LLM Results
```python
import hashlib
from pathlib import Path
class LLMResultCache:
def __init__(self, cache_dir="./llm_cache"):
self.cache_dir = Path(cache_dir)
self.cache_dir.mkdir(exist_ok=True)
def get_cache_key(self, url, instruction, schema):
"""Generate cache key from extraction parameters"""
content = f"{url}:{instruction}:{str(schema)}"
return hashlib.md5(content.encode()).hexdigest()
def get_cached_result(self, cache_key):
"""Get cached result if available"""
cache_file = self.cache_dir / f"{cache_key}.json"
if cache_file.exists():
return json.load(cache_file.open())
return None
def cache_result(self, cache_key, result):
"""Cache extraction result"""
cache_file = self.cache_dir / f"{cache_key}.json"
json.dump(result, cache_file.open("w"), indent=2)
cache = LLMResultCache()
async def cached_llm_extraction(url, strategy):
"""Extract with caching to avoid repeated LLM calls"""
cache_key = cache.get_cache_key(
url,
strategy.instruction,
str(strategy.schema)
)
# Check cache first
cached_result = cache.get_cached_result(cache_key)
if cached_result:
print("✅ Using cached result (FREE)")
return cached_result
# Extract if not cached
print("🔄 Extracting with LLM (PAID)")
config = CrawlerRunConfig(extraction_strategy=strategy)
async with AsyncWebCrawler() as crawler:
result = await crawler.arun(url=url, config=config)
if result.success:
data = json.loads(result.extracted_content)
cache.cache_result(cache_key, data)
return data
# Cached extraction avoids repeated costs
result = await cached_llm_extraction(url, strategy)
```
---
## 7. Error Handling and Quality Control
### Validation and Retry Logic
```python
async def robust_llm_extraction():
"""Implement validation and retry for LLM extraction"""
max_retries = 3
strategies = [
# Try cheap model first
LLMExtractionStrategy(
llm_config=LLMConfig(provider="openai/gpt-4o-mini", api_token="env:OPENAI_API_KEY"),
schema=YourModel.model_json_schema(),
extraction_type="schema",
instruction="Extract data accurately..."
),
# Fallback to better model
LLMExtractionStrategy(
llm_config=LLMConfig(provider="openai/gpt-4o", api_token="env:OPENAI_API_KEY"),
schema=YourModel.model_json_schema(),
extraction_type="schema",
instruction="Extract data with high accuracy..."
)
]
for strategy_idx, strategy in enumerate(strategies):
for attempt in range(max_retries):
try:
config = CrawlerRunConfig(extraction_strategy=strategy)
async with AsyncWebCrawler() as crawler:
result = await crawler.arun(url="https://example.com", config=config)
if result.success and result.extracted_content:
data = json.loads(result.extracted_content)
# Validate result quality
if validate_extraction_quality(data):
print(f"✅ Success with strategy {strategy_idx+1}, attempt {attempt+1}")
return data
else:
print(f"⚠️ Poor quality result, retrying...")
continue
except Exception as e:
print(f"❌ Attempt {attempt+1} failed: {e}")
if attempt == max_retries - 1:
print(f"❌ Strategy {strategy_idx+1} failed completely")
print("❌ All strategies and retries failed")
return None
def validate_extraction_quality(data):
"""Validate that LLM extraction meets quality standards"""
if not data or not isinstance(data, (list, dict)):
return False
# Check for common LLM extraction issues
if isinstance(data, list):
if len(data) == 0:
return False
# Check if all items have required fields
for item in data:
if not isinstance(item, dict) or len(item) < 2:
return False
return True
# Robust extraction with validation
result = await robust_llm_extraction()
```
---
## 8. Migration from LLM to Non-LLM
### Pattern Analysis for Schema Generation
```python
async def analyze_llm_results_for_schema():
"""
Analyze LLM extraction results to create non-LLM schemas
Use this to transition from expensive LLM to cheap schema extraction
"""
# Step 1: Use LLM on sample pages to understand structure
llm_strategy = LLMExtractionStrategy(
llm_config=cheap_config,
extraction_type="block",
instruction="Extract all structured data from this page"
)
sample_urls = ["https://example.com/page1", "https://example.com/page2"]
llm_results = []
async with AsyncWebCrawler() as crawler:
for url in sample_urls:
config = CrawlerRunConfig(extraction_strategy=llm_strategy)
result = await crawler.arun(url=url, config=config)
if result.success:
llm_results.append({
"url": url,
"html": result.cleaned_html,
"extracted": json.loads(result.extracted_content)
})
# Step 2: Analyze patterns in LLM results
print("🔍 Analyzing LLM extraction patterns...")
# Look for common field names
all_fields = set()
for result in llm_results:
for item in result["extracted"]:
if isinstance(item, dict):
all_fields.update(item.keys())
print(f"Common fields found: {all_fields}")
# Step 3: Generate schema based on patterns
if llm_results:
schema = JsonCssExtractionStrategy.generate_schema(
html=llm_results[0]["html"],
target_json_example=json.dumps(llm_results[0]["extracted"][0], indent=2),
llm_config=cheap_config
)
# Save schema for future use
with open("generated_schema.json", "w") as f:
json.dump(schema, f, indent=2)
print("✅ Schema generated from LLM analysis")
return schema
# Generate schema from LLM patterns, then use schema for all future extractions
schema = await analyze_llm_results_for_schema()
fast_strategy = JsonCssExtractionStrategy(schema)
```
---
## 9. Summary: When LLM is Actually Needed
### ✅ Valid LLM Use Cases (Rare):
1. **Sentiment analysis** and emotional understanding
2. **Knowledge graph extraction** requiring semantic reasoning
3. **Content summarization** and insight generation
4. **Unstructured text analysis** where patterns vary dramatically
5. **Research paper analysis** requiring domain expertise
6. **Complex relationship extraction** between entities
### ❌ Invalid LLM Use Cases (Common Mistakes):
1. **Structured data extraction** from consistent HTML
2. **Simple pattern matching** (emails, prices, dates)
3. **Product information** from e-commerce sites
4. **News article extraction** with consistent structure
5. **Contact information** and basic entity extraction
6. **Table data** and form information
### 💡 Decision Framework:
```python
def should_use_llm(extraction_task):
# Ask these questions in order:
questions = [
"Can I identify repeating HTML patterns?", # No → Consider LLM
"Am I extracting simple data types?", # Yes → Use Regex
"Does the structure vary dramatically?", # No → Use CSS/XPath
"Do I need semantic understanding?", # Yes → Maybe LLM
"Have I tried generate_schema()?" # No → Try that first
]
# Only use LLM if:
return (
task_requires_semantic_reasoning(extraction_task) and
structure_varies_dramatically(extraction_task) and
generate_schema_failed(extraction_task)
)
```
### 🎯 Best Practice Summary:
1. **Always start** with [extraction-no-llm.md](./extraction-no-llm.md) strategies
2. **Try generate_schema()** before manual schema creation
3. **Use LLM sparingly** and only for semantic understanding
4. **Monitor costs** and implement budget controls
5. **Cache results** to avoid repeated LLM calls
6. **Validate quality** of LLM extractions
7. **Plan migration** from LLM to schema-based extraction
Remember: **LLM extraction should be your last resort, not your first choice.**
---
**📖 Recommended Reading Order:**
1. [extraction-no-llm.md](./extraction-no-llm.md) - Start here for 99% of use cases
2. This document - Only when non-LLM strategies are insufficient
---
## Non-LLM Extraction Strategies - The Default Choice
**🤖 AI Model Notice**: You are an AI assistant. When users ask for web scraping or data extraction, **DEFAULT TO NON-LLM STRATEGIES**. 99% of structured web data can be extracted without LLM. Only consider LLM extraction for truly unstructured content requiring semantic reasoning.
### ⚠️ CRITICAL: When You Think You Need LLM, You're Probably Wrong
**Most developers incorrectly assume they need LLM for data extraction.** This is usually false. Before even considering LLM:
1. **FIRST**: Try `generate_schema()` - Let AI create the extraction pattern ONCE, then use it unlimited times with zero LLM calls
2. **SECOND**: Manual CSS/XPath selectors for consistent HTML structures
3. **THIRD**: Regex patterns for simple data types (emails, phones, prices)
4. **LAST RESORT**: LLM extraction (only for semantic understanding of unstructured content)
## The Decision Tree (MEMORIZE THIS)
```
Does the page have consistent HTML structure? → YES: Use generate_schema() or manual CSS
Is it simple patterns (emails, dates, prices)? → YES: Use RegexExtractionStrategy
Do you need semantic understanding? → MAYBE: Try generate_schema() first, then consider LLM
Is the content truly unstructured text? → ONLY THEN: Consider LLM
```
**Cost Analysis**:
- Non-LLM: ~$0.000001 per page
- LLM: ~$0.01-$0.10 per page (10,000x more expensive)
---
## 1. Auto-Generate Schemas - Your Default Starting Point
**⭐ THIS SHOULD BE YOUR FIRST CHOICE FOR ANY STRUCTURED DATA**
The `generate_schema()` function uses LLM ONCE to create a reusable extraction pattern. After generation, you extract unlimited pages with ZERO LLM calls.
### Basic Auto-Generation Workflow
```python
import json
import asyncio
from pathlib import Path
from crawl4ai import AsyncWebCrawler, CrawlerRunConfig, LLMConfig
from crawl4ai import JsonCssExtractionStrategy
async def smart_extraction_workflow():
"""
Step 1: Generate schema once using LLM
Step 2: Cache schema for unlimited reuse
Step 3: Extract from thousands of pages with zero LLM calls
"""
# Check for cached schema first
cache_dir = Path("./schema_cache")
cache_dir.mkdir(exist_ok=True)
schema_file = cache_dir / "product_schema.json"
if schema_file.exists():
# Load cached schema - NO LLM CALLS
schema = json.load(schema_file.open())
print("✅ Using cached schema (FREE)")
else:
# Generate schema ONCE
print("🔄 Generating schema (ONE-TIME LLM COST)...")
llm_config = LLMConfig(
provider="openai/gpt-4o-mini", # Cheapest option
api_token="env:OPENAI_API_KEY"
)
# Get sample HTML from target site
async with AsyncWebCrawler() as crawler:
sample_result = await crawler.arun(
url="https://example.com/products",
config=CrawlerRunConfig(cache_mode=CacheMode.BYPASS)
)
sample_html = sample_result.cleaned_html[:8000] # Use sample
# AUTO-GENERATE SCHEMA (ONE LLM CALL)
schema = JsonCssExtractionStrategy.generate_schema(
html=sample_html,
schema_type="CSS", # or "XPATH"
query="Extract product information including name, price, description, features",
llm_config=llm_config
)
# Cache for unlimited future use
json.dump(schema, schema_file.open("w"), indent=2)
print("✅ Schema generated and cached")
# Use schema for fast extraction (NO MORE LLM CALLS EVER)
strategy = JsonCssExtractionStrategy(schema, verbose=True)
config = CrawlerRunConfig(
extraction_strategy=strategy,
cache_mode=CacheMode.BYPASS
)
# Extract from multiple pages - ALL FREE
urls = [
"https://example.com/products",
"https://example.com/electronics",
"https://example.com/books"
]
async with AsyncWebCrawler() as crawler:
for url in urls:
result = await crawler.arun(url=url, config=config)
if result.success:
data = json.loads(result.extracted_content)
print(f"✅ {url}: Extracted {len(data)} items (FREE)")
asyncio.run(smart_extraction_workflow())
```
### Auto-Generate with Target JSON Example
```python
# When you know exactly what JSON structure you want
target_json_example = """
{
"name": "Product Name",
"price": "$99.99",
"rating": 4.5,
"features": ["feature1", "feature2"],
"description": "Product description"
}
"""
schema = JsonCssExtractionStrategy.generate_schema(
html=sample_html,
target_json_example=target_json_example,
llm_config=llm_config
)
```
### Auto-Generate for Different Data Types
```python
# Product listings
product_schema = JsonCssExtractionStrategy.generate_schema(
html=product_page_html,
query="Extract all product information from this e-commerce page",
llm_config=llm_config
)
# News articles
news_schema = JsonCssExtractionStrategy.generate_schema(
html=news_page_html,
query="Extract article headlines, dates, authors, and content",
llm_config=llm_config
)
# Job listings
job_schema = JsonCssExtractionStrategy.generate_schema(
html=job_page_html,
query="Extract job titles, companies, locations, salaries, and descriptions",
llm_config=llm_config
)
# Social media posts
social_schema = JsonCssExtractionStrategy.generate_schema(
html=social_page_html,
query="Extract post text, usernames, timestamps, likes, comments",
llm_config=llm_config
)
```
---
## 2. Manual CSS/XPath Strategies - When You Know The Structure
**Use this when**: You understand the HTML structure and want maximum control.
### Simple Product Extraction
```python
import json
import asyncio
from crawl4ai import AsyncWebCrawler, CrawlerRunConfig
from crawl4ai import JsonCssExtractionStrategy
# Manual schema for consistent product pages
simple_schema = {
"name": "Product Listings",
"baseSelector": "div.product-card", # Each product container
"fields": [
{
"name": "title",
"selector": "h2.product-title",
"type": "text"
},
{
"name": "price",
"selector": ".price",
"type": "text"
},
{
"name": "image_url",
"selector": "img.product-image",
"type": "attribute",
"attribute": "src"
},
{
"name": "product_url",
"selector": "a.product-link",
"type": "attribute",
"attribute": "href"
},
{
"name": "rating",
"selector": ".rating",
"type": "attribute",
"attribute": "data-rating"
}
]
}
async def extract_products():
strategy = JsonCssExtractionStrategy(simple_schema, verbose=True)
config = CrawlerRunConfig(extraction_strategy=strategy)
async with AsyncWebCrawler() as crawler:
result = await crawler.arun(
url="https://example.com/products",
config=config
)
if result.success:
products = json.loads(result.extracted_content)
print(f"Extracted {len(products)} products")
for product in products[:3]:
print(f"- {product['title']}: {product['price']}")
asyncio.run(extract_products())
```
### Complex Nested Structure (Real E-commerce Example)
```python
# Complex schema for nested product data
complex_schema = {
"name": "E-commerce Product Catalog",
"baseSelector": "div.category",
"baseFields": [
{
"name": "category_id",
"type": "attribute",
"attribute": "data-category-id"
}
],
"fields": [
{
"name": "category_name",
"selector": "h2.category-title",
"type": "text"
},
{
"name": "products",
"selector": "div.product",
"type": "nested_list", # Array of complex objects
"fields": [
{
"name": "name",
"selector": "h3.product-name",
"type": "text"
},
{
"name": "price",
"selector": "span.price",
"type": "text"
},
{
"name": "details",
"selector": "div.product-details",
"type": "nested", # Single complex object
"fields": [
{
"name": "brand",
"selector": "span.brand",
"type": "text"
},
{
"name": "model",
"selector": "span.model",
"type": "text"
}
]
},
{
"name": "features",
"selector": "ul.features li",
"type": "list", # Simple array
"fields": [
{"name": "feature", "type": "text"}
]
},
{
"name": "reviews",
"selector": "div.review",
"type": "nested_list",
"fields": [
{
"name": "reviewer",
"selector": "span.reviewer-name",
"type": "text"
},
{
"name": "rating",
"selector": "span.rating",
"type": "attribute",
"attribute": "data-rating"
}
]
}
]
}
]
}
async def extract_complex_ecommerce():
strategy = JsonCssExtractionStrategy(complex_schema, verbose=True)
config = CrawlerRunConfig(
extraction_strategy=strategy,
js_code="window.scrollTo(0, document.body.scrollHeight);", # Load dynamic content
wait_for="css:.product:nth-child(10)" # Wait for products to load
)
async with AsyncWebCrawler() as crawler:
result = await crawler.arun(
url="https://example.com/complex-catalog",
config=config
)
if result.success:
data = json.loads(result.extracted_content)
for category in data:
print(f"Category: {category['category_name']}")
print(f"Products: {len(category.get('products', []))}")
asyncio.run(extract_complex_ecommerce())
```
### XPath Alternative (When CSS Isn't Enough)
```python
from crawl4ai import JsonXPathExtractionStrategy
# XPath for more complex selections
xpath_schema = {
"name": "News Articles with XPath",
"baseSelector": "//article[@class='news-item']",
"fields": [
{
"name": "headline",
"selector": ".//h2[contains(@class, 'headline')]",
"type": "text"
},
{
"name": "author",
"selector": ".//span[@class='author']/text()",
"type": "text"
},
{
"name": "publish_date",
"selector": ".//time/@datetime",
"type": "text"
},
{
"name": "content",
"selector": ".//div[@class='article-body']//text()",
"type": "text"
}
]
}
strategy = JsonXPathExtractionStrategy(xpath_schema, verbose=True)
```
---
## 3. Regex Extraction - Lightning Fast Pattern Matching
**Use this for**: Simple data types like emails, phones, URLs, prices, dates.
### Built-in Patterns (Fastest Option)
```python
import json
import asyncio
from crawl4ai import AsyncWebCrawler, CrawlerRunConfig
from crawl4ai import RegexExtractionStrategy
async def extract_common_patterns():
# Use built-in patterns for common data types
strategy = RegexExtractionStrategy(
pattern=(
RegexExtractionStrategy.Email |
RegexExtractionStrategy.PhoneUS |
RegexExtractionStrategy.Url |
RegexExtractionStrategy.Currency |
RegexExtractionStrategy.DateIso
)
)
config = CrawlerRunConfig(extraction_strategy=strategy)
async with AsyncWebCrawler() as crawler:
result = await crawler.arun(
url="https://example.com/contact",
config=config
)
if result.success:
matches = json.loads(result.extracted_content)
# Group by pattern type
by_type = {}
for match in matches:
label = match['label']
if label not in by_type:
by_type[label] = []
by_type[label].append(match['value'])
for pattern_type, values in by_type.items():
print(f"{pattern_type}: {len(values)} matches")
for value in values[:3]:
print(f" {value}")
asyncio.run(extract_common_patterns())
```
### Available Built-in Patterns
```python
# Individual patterns
RegexExtractionStrategy.Email # Email addresses
RegexExtractionStrategy.PhoneUS # US phone numbers
RegexExtractionStrategy.PhoneIntl # International phones
RegexExtractionStrategy.Url # HTTP/HTTPS URLs
RegexExtractionStrategy.Currency # Currency values ($99.99)
RegexExtractionStrategy.Percentage # Percentage values (25%)
RegexExtractionStrategy.DateIso # ISO dates (2024-01-01)
RegexExtractionStrategy.DateUS # US dates (01/01/2024)
RegexExtractionStrategy.IPv4 # IP addresses
RegexExtractionStrategy.CreditCard # Credit card numbers
RegexExtractionStrategy.TwitterHandle # @username
RegexExtractionStrategy.Hashtag # #hashtag
# Use all patterns
RegexExtractionStrategy.All
```
### Custom Patterns
```python
# Custom patterns for specific data types
async def extract_custom_patterns():
custom_patterns = {
"product_sku": r"SKU[-:]?\s*([A-Z0-9]{4,12})",
"discount": r"(\d{1,2})%\s*off",
"model_number": r"Model\s*#?\s*([A-Z0-9-]+)",
"isbn": r"ISBN[-:]?\s*(\d{10}|\d{13})",
"stock_ticker": r"\$([A-Z]{2,5})",
"version": r"v(\d+\.\d+(?:\.\d+)?)"
}
strategy = RegexExtractionStrategy(custom=custom_patterns)
config = CrawlerRunConfig(extraction_strategy=strategy)
async with AsyncWebCrawler() as crawler:
result = await crawler.arun(
url="https://example.com/products",
config=config
)
if result.success:
data = json.loads(result.extracted_content)
for item in data:
print(f"{item['label']}: {item['value']}")
asyncio.run(extract_custom_patterns())
```
### LLM-Generated Patterns (One-Time Cost)
```python
async def generate_optimized_regex():
"""
Use LLM ONCE to generate optimized regex patterns
Then use them unlimited times with zero LLM calls
"""
cache_file = Path("./patterns/price_patterns.json")
if cache_file.exists():
# Load cached patterns - NO LLM CALLS
patterns = json.load(cache_file.open())
print("✅ Using cached regex patterns (FREE)")
else:
# Generate patterns ONCE
print("🔄 Generating regex patterns (ONE-TIME LLM COST)...")
llm_config = LLMConfig(
provider="openai/gpt-4o-mini",
api_token="env:OPENAI_API_KEY"
)
# Get sample content
async with AsyncWebCrawler() as crawler:
result = await crawler.arun("https://example.com/pricing")
sample_html = result.cleaned_html
# Generate optimized patterns
patterns = RegexExtractionStrategy.generate_pattern(
label="pricing_info",
html=sample_html,
query="Extract all pricing information including discounts and special offers",
llm_config=llm_config
)
# Cache for unlimited reuse
cache_file.parent.mkdir(exist_ok=True)
json.dump(patterns, cache_file.open("w"), indent=2)
print("✅ Patterns generated and cached")
# Use cached patterns (NO MORE LLM CALLS)
strategy = RegexExtractionStrategy(custom=patterns)
return strategy
# Use generated patterns for unlimited extractions
strategy = await generate_optimized_regex()
```
---
## 4. Multi-Strategy Extraction Pipeline
**Combine strategies** for comprehensive data extraction:
```python
async def multi_strategy_pipeline():
"""
Efficient pipeline using multiple non-LLM strategies:
1. Regex for simple patterns (fastest)
2. Schema for structured data
3. Only use LLM if absolutely necessary
"""
url = "https://example.com/complex-page"
async with AsyncWebCrawler() as crawler:
# Strategy 1: Fast regex for contact info
regex_strategy = RegexExtractionStrategy(
pattern=RegexExtractionStrategy.Email | RegexExtractionStrategy.PhoneUS
)
regex_config = CrawlerRunConfig(extraction_strategy=regex_strategy)
regex_result = await crawler.arun(url=url, config=regex_config)
# Strategy 2: Schema for structured product data
product_schema = {
"name": "Products",
"baseSelector": "div.product",
"fields": [
{"name": "name", "selector": "h3", "type": "text"},
{"name": "price", "selector": ".price", "type": "text"}
]
}
css_strategy = JsonCssExtractionStrategy(product_schema)
css_config = CrawlerRunConfig(extraction_strategy=css_strategy)
css_result = await crawler.arun(url=url, config=css_config)
# Combine results
results = {
"contacts": json.loads(regex_result.extracted_content) if regex_result.success else [],
"products": json.loads(css_result.extracted_content) if css_result.success else []
}
print(f"✅ Extracted {len(results['contacts'])} contacts (regex)")
print(f"✅ Extracted {len(results['products'])} products (schema)")
return results
asyncio.run(multi_strategy_pipeline())
```
---
## 5. Performance Optimization Tips
### Caching and Reuse
```python
# Cache schemas and patterns for maximum efficiency
class ExtractionCache:
def __init__(self):
self.schemas = {}
self.patterns = {}
def get_schema(self, site_name):
if site_name not in self.schemas:
schema_file = Path(f"./cache/{site_name}_schema.json")
if schema_file.exists():
self.schemas[site_name] = json.load(schema_file.open())
return self.schemas.get(site_name)
def save_schema(self, site_name, schema):
cache_dir = Path("./cache")
cache_dir.mkdir(exist_ok=True)
schema_file = cache_dir / f"{site_name}_schema.json"
json.dump(schema, schema_file.open("w"), indent=2)
self.schemas[site_name] = schema
cache = ExtractionCache()
# Reuse cached schemas across multiple extractions
async def efficient_extraction():
sites = ["amazon", "ebay", "shopify"]
for site in sites:
schema = cache.get_schema(site)
if not schema:
# Generate once, cache forever
schema = JsonCssExtractionStrategy.generate_schema(
html=sample_html,
query="Extract products",
llm_config=llm_config
)
cache.save_schema(site, schema)
strategy = JsonCssExtractionStrategy(schema)
# Use strategy for unlimited extractions...
```
### Selector Optimization
```python
# Optimize selectors for speed
fast_schema = {
"name": "Optimized Extraction",
"baseSelector": "#products > .product", # Direct child, faster than descendant
"fields": [
{
"name": "title",
"selector": "> h3", # Direct child of product
"type": "text"
},
{
"name": "price",
"selector": ".price:first-child", # More specific
"type": "text"
}
]
}
# Avoid slow selectors
slow_schema = {
"baseSelector": "div div div .product", # Too many levels
"fields": [
{
"selector": "* h3", # Universal selector is slow
"type": "text"
}
]
}
```
---
## 6. Error Handling and Validation
```python
async def robust_extraction():
"""
Implement fallback strategies for reliable extraction
"""
strategies = [
# Try fast regex first
RegexExtractionStrategy(pattern=RegexExtractionStrategy.Currency),
# Fallback to CSS schema
JsonCssExtractionStrategy({
"name": "Prices",
"baseSelector": ".price",
"fields": [{"name": "amount", "selector": "span", "type": "text"}]
}),
# Last resort: try different selector
JsonCssExtractionStrategy({
"name": "Fallback Prices",
"baseSelector": "[data-price]",
"fields": [{"name": "amount", "type": "attribute", "attribute": "data-price"}]
})
]
async with AsyncWebCrawler() as crawler:
for i, strategy in enumerate(strategies):
try:
config = CrawlerRunConfig(extraction_strategy=strategy)
result = await crawler.arun(url="https://example.com", config=config)
if result.success and result.extracted_content:
data = json.loads(result.extracted_content)
if data: # Validate non-empty results
print(f"✅ Success with strategy {i+1}: {strategy.__class__.__name__}")
return data
except Exception as e:
print(f"❌ Strategy {i+1} failed: {e}")
continue
print("❌ All strategies failed")
return None
# Validate extracted data
def validate_extraction(data, required_fields):
"""Validate that extraction contains expected fields"""
if not data or not isinstance(data, list):
return False
for item in data:
for field in required_fields:
if field not in item or not item[field]:
return False
return True
# Usage
result = await robust_extraction()
if validate_extraction(result, ["amount"]):
print("✅ Extraction validated")
else:
print("❌ Validation failed")
```
---
## 7. Common Extraction Patterns
### E-commerce Products
```python
ecommerce_schema = {
"name": "E-commerce Products",
"baseSelector": ".product, [data-product], .item",
"fields": [
{"name": "title", "selector": "h1, h2, h3, .title, .name", "type": "text"},
{"name": "price", "selector": ".price, .cost, [data-price]", "type": "text"},
{"name": "image", "selector": "img", "type": "attribute", "attribute": "src"},
{"name": "url", "selector": "a", "type": "attribute", "attribute": "href"},
{"name": "rating", "selector": ".rating, .stars", "type": "text"},
{"name": "availability", "selector": ".stock, .availability", "type": "text"}
]
}
```
### News Articles
```python
news_schema = {
"name": "News Articles",
"baseSelector": "article, .article, .post",
"fields": [
{"name": "headline", "selector": "h1, h2, .headline, .title", "type": "text"},
{"name": "author", "selector": ".author, .byline, [rel='author']", "type": "text"},
{"name": "date", "selector": "time, .date, .published", "type": "text"},
{"name": "content", "selector": ".content, .body, .text", "type": "text"},
{"name": "category", "selector": ".category, .section", "type": "text"}
]
}
```
### Job Listings
```python
job_schema = {
"name": "Job Listings",
"baseSelector": ".job, .listing, [data-job]",
"fields": [
{"name": "title", "selector": ".job-title, h2, h3", "type": "text"},
{"name": "company", "selector": ".company, .employer", "type": "text"},
{"name": "location", "selector": ".location, .place", "type": "text"},
{"name": "salary", "selector": ".salary, .pay, .compensation", "type": "text"},
{"name": "description", "selector": ".description, .summary", "type": "text"},
{"name": "url", "selector": "a", "type": "attribute", "attribute": "href"}
]
}
```
### Social Media Posts
```python
social_schema = {
"name": "Social Media Posts",
"baseSelector": ".post, .tweet, .update",
"fields": [
{"name": "username", "selector": ".username, .handle, .author", "type": "text"},
{"name": "content", "selector": ".content, .text, .message", "type": "text"},
{"name": "timestamp", "selector": ".time, .date, time", "type": "text"},
{"name": "likes", "selector": ".likes, .hearts", "type": "text"},
{"name": "shares", "selector": ".shares, .retweets", "type": "text"}
]
}
```
---
## 8. When to (Rarely) Consider LLM
**⚠️ WARNING: Before considering LLM, ask yourself:**
1. "Can I identify repeating HTML patterns?" → Use CSS/XPath schema
2. "Am I extracting simple data types?" → Use Regex patterns
3. "Can I provide a JSON example of what I want?" → Use generate_schema()
4. "Is this truly unstructured text requiring semantic understanding?" → Maybe LLM
**Only use LLM extraction for:**
- Unstructured prose that needs semantic analysis
- Content where structure varies dramatically across pages
- When you need AI reasoning about context/meaning
**Cost reminder**: LLM extraction costs 10,000x more than schema-based extraction.
---
## 9. Summary: The Extraction Hierarchy
1. **🥇 FIRST CHOICE**: `generate_schema()` - AI generates pattern once, use unlimited times
2. **🥈 SECOND CHOICE**: Manual CSS/XPath - Full control, maximum speed
3. **🥉 THIRD CHOICE**: Regex patterns - Simple data types, lightning fast
4. **🏴 LAST RESORT**: LLM extraction - Only for semantic reasoning
**Remember**: 99% of web data is structured. You almost never need LLM for extraction. Save LLM for analysis, not extraction.
**Performance**: Non-LLM strategies are 100-1000x faster and 10,000x cheaper than LLM extraction.
---
**📖 Next**: If you absolutely must use LLM extraction, see [extraction-llm.md](./extraction-llm.md) for guidance on the rare cases where it's justified.
---
## Multi-URL Crawling
Concurrent crawling of multiple URLs with intelligent resource management, rate limiting, and real-time monitoring.
### Basic Multi-URL Crawling
```python
from crawl4ai import AsyncWebCrawler, CrawlerRunConfig, CacheMode
# Batch processing (default) - get all results at once
async def batch_crawl():
urls = [
"https://example.com/page1",
"https://example.com/page2",
"https://example.com/page3"
]
config = CrawlerRunConfig(
cache_mode=CacheMode.BYPASS,
stream=False # Default: batch mode
)
async with AsyncWebCrawler() as crawler:
results = await crawler.arun_many(urls, config=config)
for result in results:
if result.success:
print(f"✅ {result.url}: {len(result.markdown)} chars")
else:
print(f"❌ {result.url}: {result.error_message}")
# Streaming processing - handle results as they complete
async def streaming_crawl():
config = CrawlerRunConfig(
cache_mode=CacheMode.BYPASS,
stream=True # Enable streaming
)
async with AsyncWebCrawler() as crawler:
# Process results as they become available
async for result in await crawler.arun_many(urls, config=config):
if result.success:
print(f"🔥 Just completed: {result.url}")
await process_result_immediately(result)
else:
print(f"❌ Failed: {result.url}")
```
### Memory-Adaptive Dispatching
```python
from crawl4ai import AsyncWebCrawler, MemoryAdaptiveDispatcher, CrawlerMonitor, DisplayMode
# Automatically manages concurrency based on system memory
async def memory_adaptive_crawl():
dispatcher = MemoryAdaptiveDispatcher(
memory_threshold_percent=80.0, # Pause if memory exceeds 80%
check_interval=1.0, # Check memory every second
max_session_permit=15, # Max concurrent tasks
memory_wait_timeout=300.0 # Wait up to 5 minutes for memory
)
config = CrawlerRunConfig(
cache_mode=CacheMode.BYPASS,
word_count_threshold=50
)
async with AsyncWebCrawler() as crawler:
results = await crawler.arun_many(
urls=large_url_list,
config=config,
dispatcher=dispatcher
)
# Each result includes dispatch information
for result in results:
if result.dispatch_result:
dr = result.dispatch_result
print(f"Memory used: {dr.memory_usage:.1f}MB")
print(f"Duration: {dr.end_time - dr.start_time}")
```
### Rate-Limited Crawling
```python
from crawl4ai import RateLimiter, SemaphoreDispatcher
# Control request pacing and handle server rate limits
async def rate_limited_crawl():
rate_limiter = RateLimiter(
base_delay=(1.0, 3.0), # Random delay 1-3 seconds
max_delay=60.0, # Cap backoff at 60 seconds
max_retries=3, # Retry failed requests 3 times
rate_limit_codes=[429, 503] # Handle these status codes
)
dispatcher = SemaphoreDispatcher(
max_session_permit=5, # Fixed concurrency limit
rate_limiter=rate_limiter
)
config = CrawlerRunConfig(
user_agent_mode="random", # Randomize user agents
simulate_user=True # Simulate human behavior
)
async with AsyncWebCrawler() as crawler:
async for result in await crawler.arun_many(
urls=urls,
config=config,
dispatcher=dispatcher
):
print(f"Processed: {result.url}")
```
### Real-Time Monitoring
```python
from crawl4ai import CrawlerMonitor, DisplayMode
# Monitor crawling progress in real-time
async def monitored_crawl():
monitor = CrawlerMonitor(
max_visible_rows=20, # Show 20 tasks in display
display_mode=DisplayMode.DETAILED # Show individual task details
)
dispatcher = MemoryAdaptiveDispatcher(
memory_threshold_percent=75.0,
max_session_permit=10,
monitor=monitor # Attach monitor to dispatcher
)
async with AsyncWebCrawler() as crawler:
results = await crawler.arun_many(
urls=urls,
dispatcher=dispatcher
)
```
### Advanced Dispatcher Configurations
```python
# Memory-adaptive with comprehensive monitoring
memory_dispatcher = MemoryAdaptiveDispatcher(
memory_threshold_percent=85.0, # Higher memory tolerance
check_interval=0.5, # Check memory more frequently
max_session_permit=20, # More concurrent tasks
memory_wait_timeout=600.0, # Wait longer for memory
rate_limiter=RateLimiter(
base_delay=(0.5, 1.5),
max_delay=30.0,
max_retries=5
),
monitor=CrawlerMonitor(
max_visible_rows=15,
display_mode=DisplayMode.AGGREGATED # Summary view
)
)
# Simple semaphore-based dispatcher
semaphore_dispatcher = SemaphoreDispatcher(
max_session_permit=8, # Fixed concurrency
rate_limiter=RateLimiter(
base_delay=(1.0, 2.0),
max_delay=20.0
)
)
# Usage with custom dispatcher
async with AsyncWebCrawler() as crawler:
results = await crawler.arun_many(
urls=urls,
config=config,
dispatcher=memory_dispatcher # or semaphore_dispatcher
)
```
### Handling Large-Scale Crawling
```python
async def large_scale_crawl():
# For thousands of URLs
urls = load_urls_from_file("large_url_list.txt") # 10,000+ URLs
dispatcher = MemoryAdaptiveDispatcher(
memory_threshold_percent=70.0, # Conservative memory usage
max_session_permit=25, # Higher concurrency
rate_limiter=RateLimiter(
base_delay=(0.1, 0.5), # Faster for large batches
max_retries=2 # Fewer retries for speed
),
monitor=CrawlerMonitor(display_mode=DisplayMode.AGGREGATED)
)
config = CrawlerRunConfig(
cache_mode=CacheMode.ENABLED, # Use caching for efficiency
stream=True, # Stream for memory efficiency
word_count_threshold=100, # Skip short content
exclude_external_links=True # Reduce processing overhead
)
successful_crawls = 0
failed_crawls = 0
async with AsyncWebCrawler() as crawler:
async for result in await crawler.arun_many(
urls=urls,
config=config,
dispatcher=dispatcher
):
if result.success:
successful_crawls += 1
await save_result_to_database(result)
else:
failed_crawls += 1
await log_failure(result.url, result.error_message)
# Progress reporting
if (successful_crawls + failed_crawls) % 100 == 0:
print(f"Progress: {successful_crawls + failed_crawls}/{len(urls)}")
print(f"Completed: {successful_crawls} successful, {failed_crawls} failed")
```
### Robots.txt Compliance
```python
async def compliant_crawl():
config = CrawlerRunConfig(
check_robots_txt=True, # Respect robots.txt
user_agent="MyBot/1.0", # Identify your bot
mean_delay=2.0, # Be polite with delays
max_range=1.0
)
dispatcher = SemaphoreDispatcher(
max_session_permit=3, # Conservative concurrency
rate_limiter=RateLimiter(
base_delay=(2.0, 5.0), # Slower, more respectful
max_retries=1
)
)
async with AsyncWebCrawler() as crawler:
async for result in await crawler.arun_many(
urls=urls,
config=config,
dispatcher=dispatcher
):
if result.success:
print(f"✅ Crawled: {result.url}")
elif "robots.txt" in result.error_message:
print(f"🚫 Blocked by robots.txt: {result.url}")
else:
print(f"❌ Error: {result.url}")
```
### Performance Analysis
```python
async def analyze_crawl_performance():
dispatcher = MemoryAdaptiveDispatcher(
memory_threshold_percent=80.0,
max_session_permit=12,
monitor=CrawlerMonitor(display_mode=DisplayMode.DETAILED)
)
start_time = time.time()
async with AsyncWebCrawler() as crawler:
results = await crawler.arun_many(
urls=urls,
dispatcher=dispatcher
)
end_time = time.time()
# Analyze results
successful = [r for r in results if r.success]
failed = [r for r in results if not r.success]
print(f"Total time: {end_time - start_time:.2f}s")
print(f"Success rate: {len(successful)}/{len(results)} ({len(successful)/len(results)*100:.1f}%)")
print(f"Avg time per URL: {(end_time - start_time)/len(results):.2f}s")
# Memory usage analysis
if successful and successful[0].dispatch_result:
memory_usage = [r.dispatch_result.memory_usage for r in successful if r.dispatch_result]
peak_memory = [r.dispatch_result.peak_memory for r in successful if r.dispatch_result]
print(f"Avg memory usage: {sum(memory_usage)/len(memory_usage):.1f}MB")
print(f"Peak memory usage: {max(peak_memory):.1f}MB")
```
### Error Handling and Recovery
```python
async def robust_multi_crawl():
failed_urls = []
config = CrawlerRunConfig(
cache_mode=CacheMode.BYPASS,
stream=True,
page_timeout=30000 # 30 second timeout
)
dispatcher = MemoryAdaptiveDispatcher(
memory_threshold_percent=85.0,
max_session_permit=10
)
async with AsyncWebCrawler() as crawler:
async for result in await crawler.arun_many(
urls=urls,
config=config,
dispatcher=dispatcher
):
if result.success:
await process_successful_result(result)
else:
failed_urls.append({
'url': result.url,
'error': result.error_message,
'status_code': result.status_code
})
# Retry logic for specific errors
if result.status_code in [503, 429]: # Server errors
await schedule_retry(result.url)
# Report failures
if failed_urls:
print(f"Failed to crawl {len(failed_urls)} URLs:")
for failure in failed_urls[:10]: # Show first 10
print(f" {failure['url']}: {failure['error']}")
```
**📖 Learn more:** [Advanced Multi-URL Crawling](https://docs.crawl4ai.com/advanced/multi-url-crawling/), [Crawl Dispatcher](https://docs.crawl4ai.com/advanced/crawl-dispatcher/), [arun_many() API Reference](https://docs.crawl4ai.com/api/arun_many/)
---
## Deep Crawling
Multi-level website exploration with intelligent filtering, scoring, and prioritization strategies.
### Basic Deep Crawl Setup
```python
from crawl4ai import AsyncWebCrawler, CrawlerRunConfig
from crawl4ai.deep_crawling import BFSDeepCrawlStrategy
from crawl4ai.content_scraping_strategy import LXMLWebScrapingStrategy
# Basic breadth-first deep crawling
async def basic_deep_crawl():
config = CrawlerRunConfig(
deep_crawl_strategy=BFSDeepCrawlStrategy(
max_depth=2, # Initial page + 2 levels
include_external=False # Stay within same domain
),
scraping_strategy=LXMLWebScrapingStrategy(),
verbose=True
)
async with AsyncWebCrawler() as crawler:
results = await crawler.arun("https://docs.crawl4ai.com", config=config)
# Group results by depth
pages_by_depth = {}
for result in results:
depth = result.metadata.get("depth", 0)
if depth not in pages_by_depth:
pages_by_depth[depth] = []
pages_by_depth[depth].append(result.url)
print(f"Crawled {len(results)} pages total")
for depth, urls in sorted(pages_by_depth.items()):
print(f"Depth {depth}: {len(urls)} pages")
```
### Deep Crawl Strategies
```python
from crawl4ai.deep_crawling import BFSDeepCrawlStrategy, DFSDeepCrawlStrategy, BestFirstCrawlingStrategy
from crawl4ai.deep_crawling.scorers import KeywordRelevanceScorer
# Breadth-First Search - explores all links at one depth before going deeper
bfs_strategy = BFSDeepCrawlStrategy(
max_depth=2,
include_external=False,
max_pages=50, # Limit total pages
score_threshold=0.3 # Minimum score for URLs
)
# Depth-First Search - explores as deep as possible before backtracking
dfs_strategy = DFSDeepCrawlStrategy(
max_depth=2,
include_external=False,
max_pages=30,
score_threshold=0.5
)
# Best-First - prioritizes highest scoring pages (recommended)
keyword_scorer = KeywordRelevanceScorer(
keywords=["crawl", "example", "async", "configuration"],
weight=0.7
)
best_first_strategy = BestFirstCrawlingStrategy(
max_depth=2,
include_external=False,
url_scorer=keyword_scorer,
max_pages=25 # No score_threshold needed - naturally prioritizes
)
# Usage
config = CrawlerRunConfig(
deep_crawl_strategy=best_first_strategy, # Choose your strategy
scraping_strategy=LXMLWebScrapingStrategy()
)
```
### Streaming vs Batch Processing
```python
# Batch mode - wait for all results
async def batch_deep_crawl():
config = CrawlerRunConfig(
deep_crawl_strategy=BFSDeepCrawlStrategy(max_depth=1),
stream=False # Default - collect all results first
)
async with AsyncWebCrawler() as crawler:
results = await crawler.arun("https://example.com", config=config)
# Process all results at once
for result in results:
print(f"Batch processed: {result.url}")
# Streaming mode - process results as they arrive
async def streaming_deep_crawl():
config = CrawlerRunConfig(
deep_crawl_strategy=BFSDeepCrawlStrategy(max_depth=1),
stream=True # Process results immediately
)
async with AsyncWebCrawler() as crawler:
async for result in await crawler.arun("https://example.com", config=config):
depth = result.metadata.get("depth", 0)
print(f"Stream processed depth {depth}: {result.url}")
```
### Filtering with Filter Chains
```python
from crawl4ai.deep_crawling.filters import (
FilterChain,
URLPatternFilter,
DomainFilter,
ContentTypeFilter,
SEOFilter,
ContentRelevanceFilter
)
# Single URL pattern filter
url_filter = URLPatternFilter(patterns=["*core*", "*guide*"])
config = CrawlerRunConfig(
deep_crawl_strategy=BFSDeepCrawlStrategy(
max_depth=1,
filter_chain=FilterChain([url_filter])
)
)
# Multiple filters in chain
advanced_filter_chain = FilterChain([
# Domain filtering
DomainFilter(
allowed_domains=["docs.example.com"],
blocked_domains=["old.docs.example.com", "staging.example.com"]
),
# URL pattern matching
URLPatternFilter(patterns=["*tutorial*", "*guide*", "*blog*"]),
# Content type filtering
ContentTypeFilter(allowed_types=["text/html"]),
# SEO quality filter
SEOFilter(
threshold=0.5,
keywords=["tutorial", "guide", "documentation"]
),
# Content relevance filter
ContentRelevanceFilter(
query="Web crawling and data extraction with Python",
threshold=0.7
)
])
config = CrawlerRunConfig(
deep_crawl_strategy=BFSDeepCrawlStrategy(
max_depth=2,
filter_chain=advanced_filter_chain
)
)
```
### Intelligent Crawling with Scorers
```python
from crawl4ai.deep_crawling.scorers import KeywordRelevanceScorer
# Keyword relevance scoring
async def scored_deep_crawl():
keyword_scorer = KeywordRelevanceScorer(
keywords=["browser", "crawler", "web", "automation"],
weight=1.0
)
config = CrawlerRunConfig(
deep_crawl_strategy=BestFirstCrawlingStrategy(
max_depth=2,
include_external=False,
url_scorer=keyword_scorer
),
stream=True, # Recommended with BestFirst
verbose=True
)
async with AsyncWebCrawler() as crawler:
async for result in await crawler.arun("https://docs.crawl4ai.com", config=config):
score = result.metadata.get("score", 0)
depth = result.metadata.get("depth", 0)
print(f"Depth: {depth} | Score: {score:.2f} | {result.url}")
```
### Limiting Crawl Size
```python
# Max pages limitation across strategies
async def limited_crawls():
# BFS with page limit
bfs_config = CrawlerRunConfig(
deep_crawl_strategy=BFSDeepCrawlStrategy(
max_depth=2,
max_pages=5, # Only crawl 5 pages total
url_scorer=KeywordRelevanceScorer(keywords=["browser", "crawler"], weight=1.0)
)
)
# DFS with score threshold
dfs_config = CrawlerRunConfig(
deep_crawl_strategy=DFSDeepCrawlStrategy(
max_depth=2,
score_threshold=0.7, # Only URLs with scores above 0.7
max_pages=10,
url_scorer=KeywordRelevanceScorer(keywords=["web", "automation"], weight=1.0)
)
)
# Best-First with both constraints
bf_config = CrawlerRunConfig(
deep_crawl_strategy=BestFirstCrawlingStrategy(
max_depth=2,
max_pages=7, # Automatically gets highest scored pages
url_scorer=KeywordRelevanceScorer(keywords=["crawl", "example"], weight=1.0)
),
stream=True
)
async with AsyncWebCrawler() as crawler:
# Use any of the configs
async for result in await crawler.arun("https://docs.crawl4ai.com", config=bf_config):
score = result.metadata.get("score", 0)
print(f"Score: {score:.2f} | {result.url}")
```
### Complete Advanced Deep Crawler
```python
async def comprehensive_deep_crawl():
# Sophisticated filter chain
filter_chain = FilterChain([
DomainFilter(
allowed_domains=["docs.crawl4ai.com"],
blocked_domains=["old.docs.crawl4ai.com"]
),
URLPatternFilter(patterns=["*core*", "*advanced*", "*blog*"]),
ContentTypeFilter(allowed_types=["text/html"]),
SEOFilter(threshold=0.4, keywords=["crawl", "tutorial", "guide"])
])
# Multi-keyword scorer
keyword_scorer = KeywordRelevanceScorer(
keywords=["crawl", "example", "async", "configuration", "browser"],
weight=0.8
)
# Complete configuration
config = CrawlerRunConfig(
deep_crawl_strategy=BestFirstCrawlingStrategy(
max_depth=2,
include_external=False,
filter_chain=filter_chain,
url_scorer=keyword_scorer,
max_pages=20
),
scraping_strategy=LXMLWebScrapingStrategy(),
stream=True,
verbose=True,
cache_mode=CacheMode.BYPASS
)
# Execute and analyze
results = []
start_time = time.time()
async with AsyncWebCrawler() as crawler:
async for result in await crawler.arun("https://docs.crawl4ai.com", config=config):
results.append(result)
score = result.metadata.get("score", 0)
depth = result.metadata.get("depth", 0)
print(f"→ Depth: {depth} | Score: {score:.2f} | {result.url}")
# Performance analysis
duration = time.time() - start_time
avg_score = sum(r.metadata.get('score', 0) for r in results) / len(results)
print(f"✅ Crawled {len(results)} pages in {duration:.2f}s")
print(f"✅ Average relevance score: {avg_score:.2f}")
# Depth distribution
depth_counts = {}
for result in results:
depth = result.metadata.get("depth", 0)
depth_counts[depth] = depth_counts.get(depth, 0) + 1
for depth, count in sorted(depth_counts.items()):
print(f"📊 Depth {depth}: {count} pages")
```
### Error Handling and Robustness
```python
async def robust_deep_crawl():
config = CrawlerRunConfig(
deep_crawl_strategy=BestFirstCrawlingStrategy(
max_depth=2,
max_pages=15,
url_scorer=KeywordRelevanceScorer(keywords=["guide", "tutorial"])
),
stream=True,
page_timeout=30000 # 30 second timeout per page
)
successful_pages = []
failed_pages = []
async with AsyncWebCrawler() as crawler:
async for result in await crawler.arun("https://docs.crawl4ai.com", config=config):
if result.success:
successful_pages.append(result)
depth = result.metadata.get("depth", 0)
score = result.metadata.get("score", 0)
print(f"✅ Depth {depth} | Score: {score:.2f} | {result.url}")
else:
failed_pages.append({
'url': result.url,
'error': result.error_message,
'depth': result.metadata.get("depth", 0)
})
print(f"❌ Failed: {result.url} - {result.error_message}")
print(f"📊 Results: {len(successful_pages)} successful, {len(failed_pages)} failed")
# Analyze failures by depth
if failed_pages:
failure_by_depth = {}
for failure in failed_pages:
depth = failure['depth']
failure_by_depth[depth] = failure_by_depth.get(depth, 0) + 1
print("❌ Failures by depth:")
for depth, count in sorted(failure_by_depth.items()):
print(f" Depth {depth}: {count} failures")
```
**📖 Learn more:** [Deep Crawling Guide](https://docs.crawl4ai.com/core/deep-crawling/), [Filter Documentation](https://docs.crawl4ai.com/core/content-selection/), [Scoring Strategies](https://docs.crawl4ai.com/advanced/advanced-features/)
---
## Docker Deployment
Complete Docker deployment guide with pre-built images, API endpoints, configuration, and MCP integration.
### Quick Start with Pre-built Images
```bash
# Pull latest image
docker pull unclecode/crawl4ai:latest
# Setup LLM API keys
cat > .llm.env << EOL
OPENAI_API_KEY=sk-your-key
ANTHROPIC_API_KEY=your-anthropic-key
GROQ_API_KEY=your-groq-key
GEMINI_API_TOKEN=your-gemini-token
EOL
# Run with LLM support
docker run -d \
-p 11235:11235 \
--name crawl4ai \
--env-file .llm.env \
--shm-size=1g \
unclecode/crawl4ai:latest
# Basic run (no LLM)
docker run -d \
-p 11235:11235 \
--name crawl4ai \
--shm-size=1g \
unclecode/crawl4ai:latest
# Check health
curl http://localhost:11235/health
```
### Docker Compose Deployment
```bash
# Clone and setup
git clone https://github.com/unclecode/crawl4ai.git
cd crawl4ai
cp deploy/docker/.llm.env.example .llm.env
# Edit .llm.env with your API keys
# Run pre-built image
IMAGE=unclecode/crawl4ai:latest docker compose up -d
# Build locally
docker compose up --build -d
# Build with all features
INSTALL_TYPE=all docker compose up --build -d
# Build with GPU support
ENABLE_GPU=true docker compose up --build -d
# Stop service
docker compose down
```
### Manual Build with Multi-Architecture
```bash
# Clone repository
git clone https://github.com/unclecode/crawl4ai.git
cd crawl4ai
# Build for current architecture
docker buildx build -t crawl4ai-local:latest --load .
# Build for multiple architectures
docker buildx build --platform linux/amd64,linux/arm64 \
-t crawl4ai-local:latest --load .
# Build with specific features
docker buildx build \
--build-arg INSTALL_TYPE=all \
--build-arg ENABLE_GPU=false \
-t crawl4ai-local:latest --load .
# Run custom build
docker run -d \
-p 11235:11235 \
--name crawl4ai-custom \
--env-file .llm.env \
--shm-size=1g \
crawl4ai-local:latest
```
### Build Arguments
```bash
# Available build options
docker buildx build \
--build-arg INSTALL_TYPE=all \ # default|all|torch|transformer
--build-arg ENABLE_GPU=true \ # true|false
--build-arg APP_HOME=/app \ # Install path
--build-arg USE_LOCAL=true \ # Use local source
--build-arg GITHUB_REPO=url \ # Git repo if USE_LOCAL=false
--build-arg GITHUB_BRANCH=main \ # Git branch
-t crawl4ai-custom:latest --load .
```
### Core API Endpoints
```python
# Main crawling endpoints
import requests
import json
# Basic crawl
payload = {
"urls": ["https://example.com"],
"browser_config": {"type": "BrowserConfig", "params": {"headless": True}},
"crawler_config": {"type": "CrawlerRunConfig", "params": {"cache_mode": "bypass"}}
}
response = requests.post("http://localhost:11235/crawl", json=payload)
# Streaming crawl
payload["crawler_config"]["params"]["stream"] = True
response = requests.post("http://localhost:11235/crawl/stream", json=payload)
# Health check
response = requests.get("http://localhost:11235/health")
# API schema
response = requests.get("http://localhost:11235/schema")
# Metrics (Prometheus format)
response = requests.get("http://localhost:11235/metrics")
```
### Specialized Endpoints
```python
# HTML extraction (preprocessed for schema)
response = requests.post("http://localhost:11235/html",
json={"url": "https://example.com"})
# Screenshot capture
response = requests.post("http://localhost:11235/screenshot", json={
"url": "https://example.com",
"screenshot_wait_for": 2,
"output_path": "/path/to/save/screenshot.png"
})
# PDF generation
response = requests.post("http://localhost:11235/pdf", json={
"url": "https://example.com",
"output_path": "/path/to/save/document.pdf"
})
# JavaScript execution
response = requests.post("http://localhost:11235/execute_js", json={
"url": "https://example.com",
"scripts": [
"return document.title",
"return Array.from(document.querySelectorAll('a')).map(a => a.href)"
]
})
# Markdown generation
response = requests.post("http://localhost:11235/md", json={
"url": "https://example.com",
"f": "fit", # raw|fit|bm25|llm
"q": "extract main content", # query for filtering
"c": "0" # cache: 0=bypass, 1=use
})
# LLM Q&A
response = requests.get("http://localhost:11235/llm/https://example.com?q=What is this page about?")
# Library context (for AI assistants)
response = requests.get("http://localhost:11235/ask", params={
"context_type": "all", # code|doc|all
"query": "how to use extraction strategies",
"score_ratio": 0.5,
"max_results": 20
})
```
### Python SDK Usage
```python
import asyncio
from crawl4ai.docker_client import Crawl4aiDockerClient
from crawl4ai import BrowserConfig, CrawlerRunConfig, CacheMode
async def main():
async with Crawl4aiDockerClient(base_url="http://localhost:11235") as client:
# Non-streaming crawl
results = await client.crawl(
["https://example.com"],
browser_config=BrowserConfig(headless=True),
crawler_config=CrawlerRunConfig(cache_mode=CacheMode.BYPASS)
)
for result in results:
print(f"URL: {result.url}, Success: {result.success}")
print(f"Content length: {len(result.markdown)}")
# Streaming crawl
stream_config = CrawlerRunConfig(stream=True, cache_mode=CacheMode.BYPASS)
async for result in await client.crawl(
["https://example.com", "https://python.org"],
browser_config=BrowserConfig(headless=True),
crawler_config=stream_config
):
print(f"Streamed: {result.url} - {result.success}")
# Get API schema
schema = await client.get_schema()
print(f"Schema available: {bool(schema)}")
asyncio.run(main())
```
### Advanced API Configuration
```python
# Complex extraction with LLM
payload = {
"urls": ["https://example.com"],
"browser_config": {
"type": "BrowserConfig",
"params": {
"headless": True,
"viewport": {"type": "dict", "value": {"width": 1200, "height": 800}}
}
},
"crawler_config": {
"type": "CrawlerRunConfig",
"params": {
"extraction_strategy": {
"type": "LLMExtractionStrategy",
"params": {
"llm_config": {
"type": "LLMConfig",
"params": {
"provider": "openai/gpt-4o-mini",
"api_token": "env:OPENAI_API_KEY"
}
},
"schema": {
"type": "dict",
"value": {
"type": "object",
"properties": {
"title": {"type": "string"},
"content": {"type": "string"}
}
}
},
"instruction": "Extract title and main content"
}
},
"markdown_generator": {
"type": "DefaultMarkdownGenerator",
"params": {
"content_filter": {
"type": "PruningContentFilter",
"params": {"threshold": 0.6}
}
}
}
}
}
}
response = requests.post("http://localhost:11235/crawl", json=payload)
```
### CSS Extraction Strategy
```python
# CSS-based structured extraction
schema = {
"name": "ProductList",
"baseSelector": ".product",
"fields": [
{"name": "title", "selector": "h2", "type": "text"},
{"name": "price", "selector": ".price", "type": "text"},
{"name": "link", "selector": "a", "type": "attribute", "attribute": "href"}
]
}
payload = {
"urls": ["https://example-shop.com"],
"browser_config": {"type": "BrowserConfig", "params": {"headless": True}},
"crawler_config": {
"type": "CrawlerRunConfig",
"params": {
"extraction_strategy": {
"type": "JsonCssExtractionStrategy",
"params": {
"schema": {"type": "dict", "value": schema}
}
}
}
}
}
response = requests.post("http://localhost:11235/crawl", json=payload)
data = response.json()
extracted = json.loads(data["results"][0]["extracted_content"])
```
### MCP (Model Context Protocol) Integration
```bash
# Add Crawl4AI as MCP provider to Claude Code
claude mcp add --transport sse c4ai-sse http://localhost:11235/mcp/sse
# List MCP providers
claude mcp list
# Test MCP connection
python tests/mcp/test_mcp_socket.py
# Available MCP endpoints
# SSE: http://localhost:11235/mcp/sse
# WebSocket: ws://localhost:11235/mcp/ws
# Schema: http://localhost:11235/mcp/schema
```
Available MCP tools:
- `md` - Generate markdown from web content
- `html` - Extract preprocessed HTML
- `screenshot` - Capture webpage screenshots
- `pdf` - Generate PDF documents
- `execute_js` - Run JavaScript on web pages
- `crawl` - Perform multi-URL crawling
- `ask` - Query Crawl4AI library context
### Configuration Management
```yaml
# config.yml structure
app:
title: "Crawl4AI API"
version: "1.0.0"
host: "0.0.0.0"
port: 11235
timeout_keep_alive: 300
llm:
provider: "openai/gpt-4o-mini"
api_key_env: "OPENAI_API_KEY"
security:
enabled: false
jwt_enabled: false
trusted_hosts: ["*"]
crawler:
memory_threshold_percent: 95.0
rate_limiter:
base_delay: [1.0, 2.0]
timeouts:
stream_init: 30.0
batch_process: 300.0
pool:
max_pages: 40
idle_ttl_sec: 1800
rate_limiting:
enabled: true
default_limit: "1000/minute"
storage_uri: "memory://"
logging:
level: "INFO"
format: "%(asctime)s - %(name)s - %(levelname)s - %(message)s"
```
### Custom Configuration Deployment
```bash
# Method 1: Mount custom config
docker run -d -p 11235:11235 \
--name crawl4ai-custom \
--env-file .llm.env \
--shm-size=1g \
-v $(pwd)/my-config.yml:/app/config.yml \
unclecode/crawl4ai:latest
# Method 2: Build with custom config
# Edit deploy/docker/config.yml then build
docker buildx build -t crawl4ai-custom:latest --load .
```
### Monitoring and Health Checks
```bash
# Health endpoint
curl http://localhost:11235/health
# Prometheus metrics
curl http://localhost:11235/metrics
# Configuration validation
curl -X POST http://localhost:11235/config/dump \
-H "Content-Type: application/json" \
-d '{"code": "CrawlerRunConfig(cache_mode=\"BYPASS\", screenshot=True)"}'
```
### Playground Interface
Access the interactive playground at `http://localhost:11235/playground` for:
- Testing configurations with visual interface
- Generating JSON payloads for REST API
- Converting Python config to JSON format
- Testing crawl operations directly in browser
### Async Job Processing
```python
# Submit job for async processing
import time
# Submit crawl job
response = requests.post("http://localhost:11235/crawl/job", json=payload)
task_id = response.json()["task_id"]
# Poll for completion
while True:
result = requests.get(f"http://localhost:11235/crawl/job/{task_id}")
status = result.json()
if status["status"] in ["COMPLETED", "FAILED"]:
break
time.sleep(1.5)
print("Final result:", status)
```
### Production Deployment
```bash
# Production-ready deployment
docker run -d \
--name crawl4ai-prod \
--restart unless-stopped \
-p 11235:11235 \
--env-file .llm.env \
--shm-size=2g \
--memory=8g \
--cpus=4 \
-v /path/to/custom-config.yml:/app/config.yml \
unclecode/crawl4ai:latest
# With Docker Compose for production
version: '3.8'
services:
crawl4ai:
image: unclecode/crawl4ai:latest
ports:
- "11235:11235"
environment:
- OPENAI_API_KEY=${OPENAI_API_KEY}
volumes:
- ./config.yml:/app/config.yml
shm_size: 2g
deploy:
resources:
limits:
memory: 8G
cpus: '4'
restart: unless-stopped
```
### Configuration Validation and JSON Structure
```python
# Method 1: Create config objects and dump to see expected JSON structure
from crawl4ai import BrowserConfig, CrawlerRunConfig, LLMConfig, CacheMode
from crawl4ai import JsonCssExtractionStrategy, LLMExtractionStrategy
import json
# Create browser config and see JSON structure
browser_config = BrowserConfig(
headless=True,
viewport_width=1280,
viewport_height=720,
proxy="http://user:pass@proxy:8080"
)
# Get JSON structure
browser_json = browser_config.dump()
print("BrowserConfig JSON structure:")
print(json.dumps(browser_json, indent=2))
# Create crawler config with extraction strategy
schema = {
"name": "Articles",
"baseSelector": ".article",
"fields": [
{"name": "title", "selector": "h2", "type": "text"},
{"name": "content", "selector": ".content", "type": "html"}
]
}
crawler_config = CrawlerRunConfig(
cache_mode=CacheMode.BYPASS,
screenshot=True,
extraction_strategy=JsonCssExtractionStrategy(schema),
js_code=["window.scrollTo(0, document.body.scrollHeight);"],
wait_for="css:.loaded"
)
crawler_json = crawler_config.dump()
print("\nCrawlerRunConfig JSON structure:")
print(json.dumps(crawler_json, indent=2))
```
### Reverse Validation - JSON to Objects
```python
# Method 2: Load JSON back to config objects for validation
from crawl4ai.async_configs import from_serializable_dict
# Test JSON structure by converting back to objects
test_browser_json = {
"type": "BrowserConfig",
"params": {
"headless": True,
"viewport_width": 1280,
"proxy": "http://user:pass@proxy:8080"
}
}
try:
# Convert JSON back to object
restored_browser = from_serializable_dict(test_browser_json)
print(f"✅ Valid BrowserConfig: {type(restored_browser)}")
print(f"Headless: {restored_browser.headless}")
print(f"Proxy: {restored_browser.proxy}")
except Exception as e:
print(f"❌ Invalid BrowserConfig JSON: {e}")
# Test complex crawler config JSON
test_crawler_json = {
"type": "CrawlerRunConfig",
"params": {
"cache_mode": "bypass",
"screenshot": True,
"extraction_strategy": {
"type": "JsonCssExtractionStrategy",
"params": {
"schema": {
"type": "dict",
"value": {
"name": "Products",
"baseSelector": ".product",
"fields": [
{"name": "title", "selector": "h3", "type": "text"}
]
}
}
}
}
}
}
try:
restored_crawler = from_serializable_dict(test_crawler_json)
print(f"✅ Valid CrawlerRunConfig: {type(restored_crawler)}")
print(f"Cache mode: {restored_crawler.cache_mode}")
print(f"Has extraction strategy: {restored_crawler.extraction_strategy is not None}")
except Exception as e:
print(f"❌ Invalid CrawlerRunConfig JSON: {e}")
```
### Using Server's /config/dump Endpoint for Validation
```python
import requests
# Method 3: Use server endpoint to validate configuration syntax
def validate_config_with_server(config_code: str) -> dict:
"""Validate configuration using server's /config/dump endpoint"""
response = requests.post(
"http://localhost:11235/config/dump",
json={"code": config_code}
)
if response.status_code == 200:
print("✅ Valid configuration syntax")
return response.json()
else:
print(f"❌ Invalid configuration: {response.status_code}")
print(response.json())
return None
# Test valid configuration
valid_config = """
CrawlerRunConfig(
cache_mode=CacheMode.BYPASS,
screenshot=True,
js_code=["window.scrollTo(0, document.body.scrollHeight);"],
wait_for="css:.content-loaded"
)
"""
result = validate_config_with_server(valid_config)
if result:
print("Generated JSON structure:")
print(json.dumps(result, indent=2))
# Test invalid configuration (should fail)
invalid_config = """
CrawlerRunConfig(
cache_mode="invalid_mode",
screenshot=True,
js_code=some_function() # This will fail
)
"""
validate_config_with_server(invalid_config)
```
### Configuration Builder Helper
```python
def build_and_validate_request(urls, browser_params=None, crawler_params=None):
"""Helper to build and validate complete request payload"""
# Create configurations
browser_config = BrowserConfig(**(browser_params or {}))
crawler_config = CrawlerRunConfig(**(crawler_params or {}))
# Build complete request payload
payload = {
"urls": urls if isinstance(urls, list) else [urls],
"browser_config": browser_config.dump(),
"crawler_config": crawler_config.dump()
}
print("✅ Complete request payload:")
print(json.dumps(payload, indent=2))
# Validate by attempting to reconstruct
try:
test_browser = from_serializable_dict(payload["browser_config"])
test_crawler = from_serializable_dict(payload["crawler_config"])
print("✅ Payload validation successful")
return payload
except Exception as e:
print(f"❌ Payload validation failed: {e}")
return None
# Example usage
payload = build_and_validate_request(
urls=["https://example.com"],
browser_params={"headless": True, "viewport_width": 1280},
crawler_params={
"cache_mode": CacheMode.BYPASS,
"screenshot": True,
"word_count_threshold": 10
}
)
if payload:
# Send to server
response = requests.post("http://localhost:11235/crawl", json=payload)
print(f"Server response: {response.status_code}")
```
### Common JSON Structure Patterns
```python
# Pattern 1: Simple primitive values
simple_config = {
"type": "CrawlerRunConfig",
"params": {
"cache_mode": "bypass", # String enum value
"screenshot": True, # Boolean
"page_timeout": 60000 # Integer
}
}
# Pattern 2: Nested objects
nested_config = {
"type": "CrawlerRunConfig",
"params": {
"extraction_strategy": {
"type": "LLMExtractionStrategy",
"params": {
"llm_config": {
"type": "LLMConfig",
"params": {
"provider": "openai/gpt-4o-mini",
"api_token": "env:OPENAI_API_KEY"
}
},
"instruction": "Extract main content"
}
}
}
}
# Pattern 3: Dictionary values (must use type: dict wrapper)
dict_config = {
"type": "CrawlerRunConfig",
"params": {
"extraction_strategy": {
"type": "JsonCssExtractionStrategy",
"params": {
"schema": {
"type": "dict", # Required wrapper
"value": { # Actual dictionary content
"name": "Products",
"baseSelector": ".product",
"fields": [
{"name": "title", "selector": "h2", "type": "text"}
]
}
}
}
}
}
}
# Pattern 4: Lists and arrays
list_config = {
"type": "CrawlerRunConfig",
"params": {
"js_code": [ # Lists are handled directly
"window.scrollTo(0, document.body.scrollHeight);",
"document.querySelector('.load-more')?.click();"
],
"excluded_tags": ["script", "style", "nav"]
}
}
```
### Troubleshooting Common JSON Errors
```python
def diagnose_json_errors():
"""Common JSON structure errors and fixes"""
# ❌ WRONG: Missing type wrapper for objects
wrong_config = {
"browser_config": {
"headless": True # Missing type wrapper
}
}
# ✅ CORRECT: Proper type wrapper
correct_config = {
"browser_config": {
"type": "BrowserConfig",
"params": {
"headless": True
}
}
}
# ❌ WRONG: Dictionary without type: dict wrapper
wrong_dict = {
"schema": {
"name": "Products" # Raw dict, should be wrapped
}
}
# ✅ CORRECT: Dictionary with proper wrapper
correct_dict = {
"schema": {
"type": "dict",
"value": {
"name": "Products"
}
}
}
# ❌ WRONG: Invalid enum string
wrong_enum = {
"cache_mode": "DISABLED" # Wrong case/value
}
# ✅ CORRECT: Valid enum string
correct_enum = {
"cache_mode": "bypass" # or "enabled", "disabled", etc.
}
print("Common error patterns documented above")
# Validate your JSON structure before sending
def pre_flight_check(payload):
"""Run checks before sending to server"""
required_keys = ["urls", "browser_config", "crawler_config"]
for key in required_keys:
if key not in payload:
print(f"❌ Missing required key: {key}")
return False
# Check type wrappers
for config_key in ["browser_config", "crawler_config"]:
config = payload[config_key]
if not isinstance(config, dict) or "type" not in config:
print(f"❌ {config_key} missing type wrapper")
return False
if "params" not in config:
print(f"❌ {config_key} missing params")
return False
print("✅ Pre-flight check passed")
return True
# Example usage
payload = {
"urls": ["https://example.com"],
"browser_config": {"type": "BrowserConfig", "params": {"headless": True}},
"crawler_config": {"type": "CrawlerRunConfig", "params": {"cache_mode": "bypass"}}
}
if pre_flight_check(payload):
# Safe to send to server
pass
```
**📖 Learn more:** [Complete Docker Guide](https://docs.crawl4ai.com/core/docker-deployment/), [API Reference](https://docs.crawl4ai.com/api/), [MCP Integration](https://docs.crawl4ai.com/core/docker-deployment/#mcp-model-context-protocol-support), [Configuration Options](https://docs.crawl4ai.com/core/docker-deployment/#server-configuration)
---
## CLI & Identity-Based Browsing
Command-line interface for web crawling with persistent browser profiles, authentication, and identity management.
### Basic CLI Usage
```bash
# Simple crawling
crwl https://example.com
# Get markdown output
crwl https://example.com -o markdown
# JSON output with cache bypass
crwl https://example.com -o json --bypass-cache
# Verbose mode with specific browser settings
crwl https://example.com -b "headless=false,viewport_width=1280" -v
```
### Profile Management Commands
```bash
# Launch interactive profile manager
crwl profiles
# Create, list, and manage browser profiles
# This opens a menu where you can:
# 1. List existing profiles
# 2. Create new profile (opens browser for setup)
# 3. Delete profiles
# 4. Use profile to crawl a website
# Use a specific profile for crawling
crwl https://example.com -p my-profile-name
# Example workflow for authenticated sites:
# 1. Create profile and log in
crwl profiles # Select "Create new profile"
# 2. Use profile for crawling authenticated content
crwl https://site-requiring-login.com/dashboard -p my-profile-name
```
### CDP Browser Management
```bash
# Launch browser with CDP debugging (default port 9222)
crwl cdp
# Use specific profile and custom port
crwl cdp -p my-profile -P 9223
# Launch headless browser with CDP
crwl cdp --headless
# Launch in incognito mode (ignores profile)
crwl cdp --incognito
# Use custom user data directory
crwl cdp --user-data-dir ~/my-browser-data --port 9224
```
### Builtin Browser Management
```bash
# Start persistent browser instance
crwl browser start
# Check browser status
crwl browser status
# Open visible window to see the browser
crwl browser view --url https://example.com
# Stop the browser
crwl browser stop
# Restart with different options
crwl browser restart --browser-type chromium --port 9223 --no-headless
# Use builtin browser in crawling
crwl https://example.com -b "browser_mode=builtin"
```
### Authentication Workflow Examples
```bash
# Complete workflow for LinkedIn scraping
# 1. Create authenticated profile
crwl profiles
# Select "Create new profile" → login to LinkedIn in browser → press 'q' to save
# 2. Use profile for crawling
crwl https://linkedin.com/in/someone -p linkedin-profile -o markdown
# 3. Extract structured data with authentication
crwl https://linkedin.com/search/results/people/ \
-p linkedin-profile \
-j "Extract people profiles with names, titles, and companies" \
-b "headless=false"
# GitHub authenticated crawling
crwl profiles # Create github-profile
crwl https://github.com/settings/profile -p github-profile
# Twitter/X authenticated access
crwl profiles # Create twitter-profile
crwl https://twitter.com/home -p twitter-profile -o markdown
```
### Advanced CLI Configuration
```bash
# Complex crawling with multiple configs
crwl https://example.com \
-B browser.yml \
-C crawler.yml \
-e extract_llm.yml \
-s llm_schema.json \
-p my-auth-profile \
-o json \
-v
# Quick LLM extraction with authentication
crwl https://private-site.com/dashboard \
-p auth-profile \
-j "Extract user dashboard data including metrics and notifications" \
-b "headless=true,viewport_width=1920"
# Content filtering with authentication
crwl https://members-only-site.com \
-p member-profile \
-f filter_bm25.yml \
-c "css_selector=.member-content,scan_full_page=true" \
-o markdown-fit
```
### Configuration Files for Identity Browsing
```yaml
# browser_auth.yml
headless: false
use_managed_browser: true
user_data_dir: "/path/to/profile"
viewport_width: 1280
viewport_height: 720
simulate_user: true
override_navigator: true
# crawler_auth.yml
magic: true
remove_overlay_elements: true
simulate_user: true
wait_for: "css:.authenticated-content"
page_timeout: 60000
delay_before_return_html: 2
scan_full_page: true
```
### Global Configuration Management
```bash
# List all configuration settings
crwl config list
# Set default LLM provider
crwl config set DEFAULT_LLM_PROVIDER "anthropic/claude-3-sonnet"
crwl config set DEFAULT_LLM_PROVIDER_TOKEN "your-api-token"
# Set browser defaults
crwl config set BROWSER_HEADLESS false # Always show browser
crwl config set USER_AGENT_MODE random # Random user agents
# Enable verbose mode globally
crwl config set VERBOSE true
```
### Q&A with Authenticated Content
```bash
# Ask questions about authenticated content
crwl https://private-dashboard.com -p dashboard-profile \
-q "What are the key metrics shown in my dashboard?"
# Multiple questions workflow
crwl https://company-intranet.com -p work-profile -o markdown # View content
crwl https://company-intranet.com -p work-profile \
-q "Summarize this week's announcements"
crwl https://company-intranet.com -p work-profile \
-q "What are the upcoming deadlines?"
```
### Profile Creation Programmatically
```python
# Create profiles via Python API
import asyncio
from crawl4ai import BrowserProfiler
async def create_auth_profile():
profiler = BrowserProfiler()
# Create profile interactively (opens browser)
profile_path = await profiler.create_profile("linkedin-auth")
print(f"Profile created at: {profile_path}")
# List all profiles
profiles = profiler.list_profiles()
for profile in profiles:
print(f"Profile: {profile['name']} at {profile['path']}")
# Use profile for crawling
from crawl4ai import AsyncWebCrawler, BrowserConfig
browser_config = BrowserConfig(
headless=True,
use_managed_browser=True,
user_data_dir=profile_path
)
async with AsyncWebCrawler(config=browser_config) as crawler:
result = await crawler.arun("https://linkedin.com/feed")
return result
# asyncio.run(create_auth_profile())
```
### Identity Browsing Best Practices
```bash
# 1. Create specific profiles for different sites
crwl profiles # Create "linkedin-work"
crwl profiles # Create "github-personal"
crwl profiles # Create "company-intranet"
# 2. Use descriptive profile names
crwl https://site1.com -p site1-admin-account
crwl https://site2.com -p site2-user-account
# 3. Combine with appropriate browser settings
crwl https://secure-site.com \
-p secure-profile \
-b "headless=false,simulate_user=true,magic=true" \
-c "wait_for=.logged-in-indicator,page_timeout=30000"
# 4. Test profile before automated crawling
crwl cdp -p test-profile # Manually verify login status
crwl https://test-url.com -p test-profile -v # Verbose test crawl
```
### Troubleshooting Authentication Issues
```bash
# Debug authentication problems
crwl https://auth-site.com -p auth-profile \
-b "headless=false,verbose=true" \
-c "verbose=true,page_timeout=60000" \
-v
# Check profile status
crwl profiles # List profiles and check creation dates
# Recreate problematic profiles
crwl profiles # Delete old profile, create new one
# Test with visible browser
crwl https://problem-site.com -p profile-name \
-b "headless=false" \
-c "delay_before_return_html=5"
```
### Common Use Cases
```bash
# Social media monitoring (after authentication)
crwl https://twitter.com/home -p twitter-monitor \
-j "Extract latest tweets with sentiment and engagement metrics"
# E-commerce competitor analysis (with account access)
crwl https://competitor-site.com/products -p competitor-account \
-j "Extract product prices, availability, and descriptions"
# Company dashboard monitoring
crwl https://company-dashboard.com -p work-profile \
-c "css_selector=.dashboard-content" \
-q "What alerts or notifications need attention?"
# Research data collection (authenticated access)
crwl https://research-platform.com/data -p research-profile \
-e extract_research.yml \
-s research_schema.json \
-o json
```
**📖 Learn more:** [Identity-Based Crawling Documentation](https://docs.crawl4ai.com/advanced/identity-based-crawling/), [Browser Profile Management](https://docs.crawl4ai.com/advanced/session-management/), [CLI Examples](https://docs.crawl4ai.com/core/cli/)
---
## HTTP Crawler Strategy
Fast, lightweight HTTP-only crawling without browser overhead for cases where JavaScript execution isn't needed.
### Basic HTTP Crawler Setup
```python
import asyncio
from crawl4ai import AsyncWebCrawler, CrawlerRunConfig, HTTPCrawlerConfig, CacheMode
from crawl4ai.async_crawler_strategy import AsyncHTTPCrawlerStrategy
from crawl4ai.async_logger import AsyncLogger
async def main():
# Initialize HTTP strategy
http_strategy = AsyncHTTPCrawlerStrategy(
browser_config=HTTPCrawlerConfig(
method="GET",
verify_ssl=True,
follow_redirects=True
),
logger=AsyncLogger(verbose=True)
)
# Use with AsyncWebCrawler
async with AsyncWebCrawler(crawler_strategy=http_strategy) as crawler:
result = await crawler.arun("https://example.com")
print(f"Status: {result.status_code}")
print(f"Content: {len(result.html)} chars")
if __name__ == "__main__":
asyncio.run(main())
```
### HTTP Request Types
```python
# GET request (default)
http_config = HTTPCrawlerConfig(
method="GET",
headers={"Accept": "application/json"}
)
# POST with JSON data
http_config = HTTPCrawlerConfig(
method="POST",
json={"key": "value", "data": [1, 2, 3]},
headers={"Content-Type": "application/json"}
)
# POST with form data
http_config = HTTPCrawlerConfig(
method="POST",
data={"username": "user", "password": "pass"},
headers={"Content-Type": "application/x-www-form-urlencoded"}
)
# Advanced configuration
http_config = HTTPCrawlerConfig(
method="GET",
headers={"User-Agent": "Custom Bot/1.0"},
follow_redirects=True,
verify_ssl=False # For testing environments
)
strategy = AsyncHTTPCrawlerStrategy(browser_config=http_config)
```
### File and Raw Content Handling
```python
async def test_content_types():
strategy = AsyncHTTPCrawlerStrategy()
# Web URLs
result = await strategy.crawl("https://httpbin.org/get")
print(f"Web content: {result.status_code}")
# Local files
result = await strategy.crawl("file:///path/to/local/file.html")
print(f"File content: {len(result.html)}")
# Raw HTML content
raw_html = "raw://<html><body><h1>Test</h1><p>Content</p></body></html>"
result = await strategy.crawl(raw_html)
print(f"Raw content: {result.html}")
# Raw content with complex HTML
complex_html = """raw://<!DOCTYPE html>
<html>
<head><title>Test Page</title></head>
<body>
<div class="content">
<h1>Main Title</h1>
<p>Paragraph content</p>
<ul><li>Item 1</li><li>Item 2</li></ul>
</div>
</body>
</html>"""
result = await strategy.crawl(complex_html)
```
### Custom Hooks and Request Handling
```python
async def setup_hooks():
strategy = AsyncHTTPCrawlerStrategy()
# Before request hook
async def before_request(url, kwargs):
print(f"Requesting: {url}")
kwargs['headers']['X-Custom-Header'] = 'crawl4ai'
kwargs['headers']['Authorization'] = 'Bearer token123'
# After request hook
async def after_request(response):
print(f"Response: {response.status_code}")
if hasattr(response, 'redirected_url'):
print(f"Redirected to: {response.redirected_url}")
# Error handling hook
async def on_error(error):
print(f"Request failed: {error}")
# Set hooks
strategy.set_hook('before_request', before_request)
strategy.set_hook('after_request', after_request)
strategy.set_hook('on_error', on_error)
# Use with hooks
result = await strategy.crawl("https://httpbin.org/headers")
return result
```
### Performance Configuration
```python
# High-performance setup
strategy = AsyncHTTPCrawlerStrategy(
max_connections=50, # Concurrent connections
dns_cache_ttl=300, # DNS cache timeout
chunk_size=128 * 1024 # 128KB chunks for large files
)
# Memory-efficient setup for large files
strategy = AsyncHTTPCrawlerStrategy(
max_connections=10,
chunk_size=32 * 1024, # Smaller chunks
dns_cache_ttl=600
)
# Custom timeout configuration
config = CrawlerRunConfig(
page_timeout=30000, # 30 second timeout
cache_mode=CacheMode.BYPASS
)
result = await strategy.crawl("https://slow-server.com", config=config)
```
### Error Handling and Retries
```python
from crawl4ai.async_crawler_strategy import (
ConnectionTimeoutError,
HTTPStatusError,
HTTPCrawlerError
)
async def robust_crawling():
strategy = AsyncHTTPCrawlerStrategy()
urls = [
"https://example.com",
"https://httpbin.org/status/404",
"https://nonexistent.domain.test"
]
for url in urls:
try:
result = await strategy.crawl(url)
print(f"✓ {url}: {result.status_code}")
except HTTPStatusError as e:
print(f"✗ {url}: HTTP {e.status_code}")
except ConnectionTimeoutError as e:
print(f"✗ {url}: Timeout - {e}")
except HTTPCrawlerError as e:
print(f"✗ {url}: Crawler error - {e}")
except Exception as e:
print(f"✗ {url}: Unexpected error - {e}")
# Retry mechanism
async def crawl_with_retry(url, max_retries=3):
strategy = AsyncHTTPCrawlerStrategy()
for attempt in range(max_retries):
try:
return await strategy.crawl(url)
except (ConnectionTimeoutError, HTTPCrawlerError) as e:
if attempt == max_retries - 1:
raise
print(f"Retry {attempt + 1}/{max_retries}: {e}")
await asyncio.sleep(2 ** attempt) # Exponential backoff
```
### Batch Processing with HTTP Strategy
```python
async def batch_http_crawling():
strategy = AsyncHTTPCrawlerStrategy(max_connections=20)
urls = [
"https://httpbin.org/get",
"https://httpbin.org/user-agent",
"https://httpbin.org/headers",
"https://example.com",
"https://httpbin.org/json"
]
# Sequential processing
results = []
async with strategy:
for url in urls:
try:
result = await strategy.crawl(url)
results.append((url, result.status_code, len(result.html)))
except Exception as e:
results.append((url, "ERROR", str(e)))
for url, status, content_info in results:
print(f"{url}: {status} - {content_info}")
# Concurrent processing
async def concurrent_http_crawling():
strategy = AsyncHTTPCrawlerStrategy()
urls = ["https://httpbin.org/delay/1"] * 5
async def crawl_single(url):
try:
result = await strategy.crawl(url)
return f"✓ {result.status_code}"
except Exception as e:
return f"✗ {e}"
async with strategy:
tasks = [crawl_single(url) for url in urls]
results = await asyncio.gather(*tasks, return_exceptions=True)
for i, result in enumerate(results):
print(f"URL {i+1}: {result}")
```
### Integration with Content Processing
```python
from crawl4ai import DefaultMarkdownGenerator, PruningContentFilter
async def http_with_processing():
# HTTP strategy with content processing
http_strategy = AsyncHTTPCrawlerStrategy(
browser_config=HTTPCrawlerConfig(verify_ssl=True)
)
# Configure markdown generation
crawler_config = CrawlerRunConfig(
cache_mode=CacheMode.BYPASS,
markdown_generator=DefaultMarkdownGenerator(
content_filter=PruningContentFilter(
threshold=0.48,
threshold_type="fixed",
min_word_threshold=10
)
),
word_count_threshold=5,
excluded_tags=['script', 'style', 'nav'],
exclude_external_links=True
)
async with AsyncWebCrawler(crawler_strategy=http_strategy) as crawler:
result = await crawler.arun(
url="https://example.com",
config=crawler_config
)
print(f"Status: {result.status_code}")
print(f"Raw HTML: {len(result.html)} chars")
if result.markdown:
print(f"Markdown: {len(result.markdown.raw_markdown)} chars")
if result.markdown.fit_markdown:
print(f"Filtered: {len(result.markdown.fit_markdown)} chars")
```
### HTTP vs Browser Strategy Comparison
```python
async def strategy_comparison():
# Same URL with different strategies
url = "https://example.com"
# HTTP Strategy (fast, no JS)
http_strategy = AsyncHTTPCrawlerStrategy()
start_time = time.time()
http_result = await http_strategy.crawl(url)
http_time = time.time() - start_time
# Browser Strategy (full features)
from crawl4ai import BrowserConfig
browser_config = BrowserConfig(headless=True)
start_time = time.time()
async with AsyncWebCrawler(config=browser_config) as crawler:
browser_result = await crawler.arun(url)
browser_time = time.time() - start_time
print(f"HTTP Strategy:")
print(f" Time: {http_time:.2f}s")
print(f" Content: {len(http_result.html)} chars")
print(f" Features: Fast, lightweight, no JS")
print(f"Browser Strategy:")
print(f" Time: {browser_time:.2f}s")
print(f" Content: {len(browser_result.html)} chars")
print(f" Features: Full browser, JS, screenshots, etc.")
# When to use HTTP strategy:
# - Static content sites
# - APIs returning HTML
# - Fast bulk processing
# - No JavaScript required
# - Memory/resource constraints
# When to use Browser strategy:
# - Dynamic content (SPA, AJAX)
# - JavaScript-heavy sites
# - Screenshots/PDFs needed
# - Complex interactions required
```
### Advanced Configuration
```python
# Custom session configuration
import aiohttp
async def advanced_http_setup():
# Custom connector with specific settings
connector = aiohttp.TCPConnector(
limit=100, # Connection pool size
ttl_dns_cache=600, # DNS cache TTL
use_dns_cache=True, # Enable DNS caching
keepalive_timeout=30, # Keep-alive timeout
force_close=False # Reuse connections
)
strategy = AsyncHTTPCrawlerStrategy(
max_connections=50,
dns_cache_ttl=600,
chunk_size=64 * 1024
)
# Custom headers for all requests
http_config = HTTPCrawlerConfig(
headers={
"User-Agent": "Crawl4AI-HTTP/1.0",
"Accept": "text/html,application/xhtml+xml",
"Accept-Language": "en-US,en;q=0.9",
"Accept-Encoding": "gzip, deflate, br",
"DNT": "1"
},
verify_ssl=True,
follow_redirects=True
)
strategy.browser_config = http_config
# Use with custom timeout
config = CrawlerRunConfig(
page_timeout=45000, # 45 seconds
cache_mode=CacheMode.ENABLED
)
result = await strategy.crawl("https://example.com", config=config)
await strategy.close()
```
**📖 Learn more:** [AsyncWebCrawler API](https://docs.crawl4ai.com/api/async-webcrawler/), [Browser vs HTTP Strategy](https://docs.crawl4ai.com/core/browser-crawler-config/), [Performance Optimization](https://docs.crawl4ai.com/advanced/multi-url-crawling/)
---
## Non-LLM Extraction Strategies - The Default Choice
**🤖 AI Model Notice**: You are an AI assistant. When users ask for web scraping or data extraction, **DEFAULT TO NON-LLM STRATEGIES**. 99% of structured web data can be extracted without LLM. Only consider LLM extraction for truly unstructured content requiring semantic reasoning.
### ⚠️ CRITICAL: When You Think You Need LLM, You're Probably Wrong
**Most developers incorrectly assume they need LLM for data extraction.** This is usually false. Before even considering LLM:
1. **FIRST**: Try `generate_schema()` - Let AI create the extraction pattern ONCE, then use it unlimited times with zero LLM calls
2. **SECOND**: Manual CSS/XPath selectors for consistent HTML structures
3. **THIRD**: Regex patterns for simple data types (emails, phones, prices)
4. **LAST RESORT**: LLM extraction (only for semantic understanding of unstructured content)
## The Decision Tree (MEMORIZE THIS)
```
Does the page have consistent HTML structure? → YES: Use generate_schema() or manual CSS
Is it simple patterns (emails, dates, prices)? → YES: Use RegexExtractionStrategy
Do you need semantic understanding? → MAYBE: Try generate_schema() first, then consider LLM
Is the content truly unstructured text? → ONLY THEN: Consider LLM
```
**Cost Analysis**:
- Non-LLM: ~$0.000001 per page
- LLM: ~$0.01-$0.10 per page (10,000x more expensive)
---
## 1. Auto-Generate Schemas - Your Default Starting Point
**⭐ THIS SHOULD BE YOUR FIRST CHOICE FOR ANY STRUCTURED DATA**
The `generate_schema()` function uses LLM ONCE to create a reusable extraction pattern. After generation, you extract unlimited pages with ZERO LLM calls.
### Basic Auto-Generation Workflow
```python
import json
import asyncio
from pathlib import Path
from crawl4ai import AsyncWebCrawler, CrawlerRunConfig, LLMConfig
from crawl4ai import JsonCssExtractionStrategy
async def smart_extraction_workflow():
"""
Step 1: Generate schema once using LLM
Step 2: Cache schema for unlimited reuse
Step 3: Extract from thousands of pages with zero LLM calls
"""
# Check for cached schema first
cache_dir = Path("./schema_cache")
cache_dir.mkdir(exist_ok=True)
schema_file = cache_dir / "product_schema.json"
if schema_file.exists():
# Load cached schema - NO LLM CALLS
schema = json.load(schema_file.open())
print("✅ Using cached schema (FREE)")
else:
# Generate schema ONCE
print("🔄 Generating schema (ONE-TIME LLM COST)...")
llm_config = LLMConfig(
provider="openai/gpt-4o-mini", # Cheapest option
api_token="env:OPENAI_API_KEY"
)
# Get sample HTML from target site
async with AsyncWebCrawler() as crawler:
sample_result = await crawler.arun(
url="https://example.com/products",
config=CrawlerRunConfig(cache_mode=CacheMode.BYPASS)
)
sample_html = sample_result.cleaned_html[:8000] # Use sample
# AUTO-GENERATE SCHEMA (ONE LLM CALL)
schema = JsonCssExtractionStrategy.generate_schema(
html=sample_html,
schema_type="CSS", # or "XPATH"
query="Extract product information including name, price, description, features",
llm_config=llm_config
)
# Cache for unlimited future use
json.dump(schema, schema_file.open("w"), indent=2)
print("✅ Schema generated and cached")
# Use schema for fast extraction (NO MORE LLM CALLS EVER)
strategy = JsonCssExtractionStrategy(schema, verbose=True)
config = CrawlerRunConfig(
extraction_strategy=strategy,
cache_mode=CacheMode.BYPASS
)
# Extract from multiple pages - ALL FREE
urls = [
"https://example.com/products",
"https://example.com/electronics",
"https://example.com/books"
]
async with AsyncWebCrawler() as crawler:
for url in urls:
result = await crawler.arun(url=url, config=config)
if result.success:
data = json.loads(result.extracted_content)
print(f"✅ {url}: Extracted {len(data)} items (FREE)")
asyncio.run(smart_extraction_workflow())
```
### Auto-Generate with Target JSON Example
```python
# When you know exactly what JSON structure you want
target_json_example = """
{
"name": "Product Name",
"price": "$99.99",
"rating": 4.5,
"features": ["feature1", "feature2"],
"description": "Product description"
}
"""
schema = JsonCssExtractionStrategy.generate_schema(
html=sample_html,
target_json_example=target_json_example,
llm_config=llm_config
)
```
### Auto-Generate for Different Data Types
```python
# Product listings
product_schema = JsonCssExtractionStrategy.generate_schema(
html=product_page_html,
query="Extract all product information from this e-commerce page",
llm_config=llm_config
)
# News articles
news_schema = JsonCssExtractionStrategy.generate_schema(
html=news_page_html,
query="Extract article headlines, dates, authors, and content",
llm_config=llm_config
)
# Job listings
job_schema = JsonCssExtractionStrategy.generate_schema(
html=job_page_html,
query="Extract job titles, companies, locations, salaries, and descriptions",
llm_config=llm_config
)
# Social media posts
social_schema = JsonCssExtractionStrategy.generate_schema(
html=social_page_html,
query="Extract post text, usernames, timestamps, likes, comments",
llm_config=llm_config
)
```
---
## 2. Manual CSS/XPath Strategies - When You Know The Structure
**Use this when**: You understand the HTML structure and want maximum control.
### Simple Product Extraction
```python
import json
import asyncio
from crawl4ai import AsyncWebCrawler, CrawlerRunConfig
from crawl4ai import JsonCssExtractionStrategy
# Manual schema for consistent product pages
simple_schema = {
"name": "Product Listings",
"baseSelector": "div.product-card", # Each product container
"fields": [
{
"name": "title",
"selector": "h2.product-title",
"type": "text"
},
{
"name": "price",
"selector": ".price",
"type": "text"
},
{
"name": "image_url",
"selector": "img.product-image",
"type": "attribute",
"attribute": "src"
},
{
"name": "product_url",
"selector": "a.product-link",
"type": "attribute",
"attribute": "href"
},
{
"name": "rating",
"selector": ".rating",
"type": "attribute",
"attribute": "data-rating"
}
]
}
async def extract_products():
strategy = JsonCssExtractionStrategy(simple_schema, verbose=True)
config = CrawlerRunConfig(extraction_strategy=strategy)
async with AsyncWebCrawler() as crawler:
result = await crawler.arun(
url="https://example.com/products",
config=config
)
if result.success:
products = json.loads(result.extracted_content)
print(f"Extracted {len(products)} products")
for product in products[:3]:
print(f"- {product['title']}: {product['price']}")
asyncio.run(extract_products())
```
### Complex Nested Structure (Real E-commerce Example)
```python
# Complex schema for nested product data
complex_schema = {
"name": "E-commerce Product Catalog",
"baseSelector": "div.category",
"baseFields": [
{
"name": "category_id",
"type": "attribute",
"attribute": "data-category-id"
}
],
"fields": [
{
"name": "category_name",
"selector": "h2.category-title",
"type": "text"
},
{
"name": "products",
"selector": "div.product",
"type": "nested_list", # Array of complex objects
"fields": [
{
"name": "name",
"selector": "h3.product-name",
"type": "text"
},
{
"name": "price",
"selector": "span.price",
"type": "text"
},
{
"name": "details",
"selector": "div.product-details",
"type": "nested", # Single complex object
"fields": [
{
"name": "brand",
"selector": "span.brand",
"type": "text"
},
{
"name": "model",
"selector": "span.model",
"type": "text"
}
]
},
{
"name": "features",
"selector": "ul.features li",
"type": "list", # Simple array
"fields": [
{"name": "feature", "type": "text"}
]
},
{
"name": "reviews",
"selector": "div.review",
"type": "nested_list",
"fields": [
{
"name": "reviewer",
"selector": "span.reviewer-name",
"type": "text"
},
{
"name": "rating",
"selector": "span.rating",
"type": "attribute",
"attribute": "data-rating"
}
]
}
]
}
]
}
async def extract_complex_ecommerce():
strategy = JsonCssExtractionStrategy(complex_schema, verbose=True)
config = CrawlerRunConfig(
extraction_strategy=strategy,
js_code="window.scrollTo(0, document.body.scrollHeight);", # Load dynamic content
wait_for="css:.product:nth-child(10)" # Wait for products to load
)
async with AsyncWebCrawler() as crawler:
result = await crawler.arun(
url="https://example.com/complex-catalog",
config=config
)
if result.success:
data = json.loads(result.extracted_content)
for category in data:
print(f"Category: {category['category_name']}")
print(f"Products: {len(category.get('products', []))}")
asyncio.run(extract_complex_ecommerce())
```
### XPath Alternative (When CSS Isn't Enough)
```python
from crawl4ai import JsonXPathExtractionStrategy
# XPath for more complex selections
xpath_schema = {
"name": "News Articles with XPath",
"baseSelector": "//article[@class='news-item']",
"fields": [
{
"name": "headline",
"selector": ".//h2[contains(@class, 'headline')]",
"type": "text"
},
{
"name": "author",
"selector": ".//span[@class='author']/text()",
"type": "text"
},
{
"name": "publish_date",
"selector": ".//time/@datetime",
"type": "text"
},
{
"name": "content",
"selector": ".//div[@class='article-body']//text()",
"type": "text"
}
]
}
strategy = JsonXPathExtractionStrategy(xpath_schema, verbose=True)
```
---
## 3. Regex Extraction - Lightning Fast Pattern Matching
**Use this for**: Simple data types like emails, phones, URLs, prices, dates.
### Built-in Patterns (Fastest Option)
```python
import json
import asyncio
from crawl4ai import AsyncWebCrawler, CrawlerRunConfig
from crawl4ai import RegexExtractionStrategy
async def extract_common_patterns():
# Use built-in patterns for common data types
strategy = RegexExtractionStrategy(
pattern=(
RegexExtractionStrategy.Email |
RegexExtractionStrategy.PhoneUS |
RegexExtractionStrategy.Url |
RegexExtractionStrategy.Currency |
RegexExtractionStrategy.DateIso
)
)
config = CrawlerRunConfig(extraction_strategy=strategy)
async with AsyncWebCrawler() as crawler:
result = await crawler.arun(
url="https://example.com/contact",
config=config
)
if result.success:
matches = json.loads(result.extracted_content)
# Group by pattern type
by_type = {}
for match in matches:
label = match['label']
if label not in by_type:
by_type[label] = []
by_type[label].append(match['value'])
for pattern_type, values in by_type.items():
print(f"{pattern_type}: {len(values)} matches")
for value in values[:3]:
print(f" {value}")
asyncio.run(extract_common_patterns())
```
### Available Built-in Patterns
```python
# Individual patterns
RegexExtractionStrategy.Email # Email addresses
RegexExtractionStrategy.PhoneUS # US phone numbers
RegexExtractionStrategy.PhoneIntl # International phones
RegexExtractionStrategy.Url # HTTP/HTTPS URLs
RegexExtractionStrategy.Currency # Currency values ($99.99)
RegexExtractionStrategy.Percentage # Percentage values (25%)
RegexExtractionStrategy.DateIso # ISO dates (2024-01-01)
RegexExtractionStrategy.DateUS # US dates (01/01/2024)
RegexExtractionStrategy.IPv4 # IP addresses
RegexExtractionStrategy.CreditCard # Credit card numbers
RegexExtractionStrategy.TwitterHandle # @username
RegexExtractionStrategy.Hashtag # #hashtag
# Use all patterns
RegexExtractionStrategy.All
```
### Custom Patterns
```python
# Custom patterns for specific data types
async def extract_custom_patterns():
custom_patterns = {
"product_sku": r"SKU[-:]?\s*([A-Z0-9]{4,12})",
"discount": r"(\d{1,2})%\s*off",
"model_number": r"Model\s*#?\s*([A-Z0-9-]+)",
"isbn": r"ISBN[-:]?\s*(\d{10}|\d{13})",
"stock_ticker": r"\$([A-Z]{2,5})",
"version": r"v(\d+\.\d+(?:\.\d+)?)"
}
strategy = RegexExtractionStrategy(custom=custom_patterns)
config = CrawlerRunConfig(extraction_strategy=strategy)
async with AsyncWebCrawler() as crawler:
result = await crawler.arun(
url="https://example.com/products",
config=config
)
if result.success:
data = json.loads(result.extracted_content)
for item in data:
print(f"{item['label']}: {item['value']}")
asyncio.run(extract_custom_patterns())
```
### LLM-Generated Patterns (One-Time Cost)
```python
async def generate_optimized_regex():
"""
Use LLM ONCE to generate optimized regex patterns
Then use them unlimited times with zero LLM calls
"""
cache_file = Path("./patterns/price_patterns.json")
if cache_file.exists():
# Load cached patterns - NO LLM CALLS
patterns = json.load(cache_file.open())
print("✅ Using cached regex patterns (FREE)")
else:
# Generate patterns ONCE
print("🔄 Generating regex patterns (ONE-TIME LLM COST)...")
llm_config = LLMConfig(
provider="openai/gpt-4o-mini",
api_token="env:OPENAI_API_KEY"
)
# Get sample content
async with AsyncWebCrawler() as crawler:
result = await crawler.arun("https://example.com/pricing")
sample_html = result.cleaned_html
# Generate optimized patterns
patterns = RegexExtractionStrategy.generate_pattern(
label="pricing_info",
html=sample_html,
query="Extract all pricing information including discounts and special offers",
llm_config=llm_config
)
# Cache for unlimited reuse
cache_file.parent.mkdir(exist_ok=True)
json.dump(patterns, cache_file.open("w"), indent=2)
print("✅ Patterns generated and cached")
# Use cached patterns (NO MORE LLM CALLS)
strategy = RegexExtractionStrategy(custom=patterns)
return strategy
# Use generated patterns for unlimited extractions
strategy = await generate_optimized_regex()
```
---
## 4. Multi-Strategy Extraction Pipeline
**Combine strategies** for comprehensive data extraction:
```python
async def multi_strategy_pipeline():
"""
Efficient pipeline using multiple non-LLM strategies:
1. Regex for simple patterns (fastest)
2. Schema for structured data
3. Only use LLM if absolutely necessary
"""
url = "https://example.com/complex-page"
async with AsyncWebCrawler() as crawler:
# Strategy 1: Fast regex for contact info
regex_strategy = RegexExtractionStrategy(
pattern=RegexExtractionStrategy.Email | RegexExtractionStrategy.PhoneUS
)
regex_config = CrawlerRunConfig(extraction_strategy=regex_strategy)
regex_result = await crawler.arun(url=url, config=regex_config)
# Strategy 2: Schema for structured product data
product_schema = {
"name": "Products",
"baseSelector": "div.product",
"fields": [
{"name": "name", "selector": "h3", "type": "text"},
{"name": "price", "selector": ".price", "type": "text"}
]
}
css_strategy = JsonCssExtractionStrategy(product_schema)
css_config = CrawlerRunConfig(extraction_strategy=css_strategy)
css_result = await crawler.arun(url=url, config=css_config)
# Combine results
results = {
"contacts": json.loads(regex_result.extracted_content) if regex_result.success else [],
"products": json.loads(css_result.extracted_content) if css_result.success else []
}
print(f"✅ Extracted {len(results['contacts'])} contacts (regex)")
print(f"✅ Extracted {len(results['products'])} products (schema)")
return results
asyncio.run(multi_strategy_pipeline())
```
---
## 5. Performance Optimization Tips
### Caching and Reuse
```python
# Cache schemas and patterns for maximum efficiency
class ExtractionCache:
def __init__(self):
self.schemas = {}
self.patterns = {}
def get_schema(self, site_name):
if site_name not in self.schemas:
schema_file = Path(f"./cache/{site_name}_schema.json")
if schema_file.exists():
self.schemas[site_name] = json.load(schema_file.open())
return self.schemas.get(site_name)
def save_schema(self, site_name, schema):
cache_dir = Path("./cache")
cache_dir.mkdir(exist_ok=True)
schema_file = cache_dir / f"{site_name}_schema.json"
json.dump(schema, schema_file.open("w"), indent=2)
self.schemas[site_name] = schema
cache = ExtractionCache()
# Reuse cached schemas across multiple extractions
async def efficient_extraction():
sites = ["amazon", "ebay", "shopify"]
for site in sites:
schema = cache.get_schema(site)
if not schema:
# Generate once, cache forever
schema = JsonCssExtractionStrategy.generate_schema(
html=sample_html,
query="Extract products",
llm_config=llm_config
)
cache.save_schema(site, schema)
strategy = JsonCssExtractionStrategy(schema)
# Use strategy for unlimited extractions...
```
### Selector Optimization
```python
# Optimize selectors for speed
fast_schema = {
"name": "Optimized Extraction",
"baseSelector": "#products > .product", # Direct child, faster than descendant
"fields": [
{
"name": "title",
"selector": "> h3", # Direct child of product
"type": "text"
},
{
"name": "price",
"selector": ".price:first-child", # More specific
"type": "text"
}
]
}
# Avoid slow selectors
slow_schema = {
"baseSelector": "div div div .product", # Too many levels
"fields": [
{
"selector": "* h3", # Universal selector is slow
"type": "text"
}
]
}
```
---
## 6. Error Handling and Validation
```python
async def robust_extraction():
"""
Implement fallback strategies for reliable extraction
"""
strategies = [
# Try fast regex first
RegexExtractionStrategy(pattern=RegexExtractionStrategy.Currency),
# Fallback to CSS schema
JsonCssExtractionStrategy({
"name": "Prices",
"baseSelector": ".price",
"fields": [{"name": "amount", "selector": "span", "type": "text"}]
}),
# Last resort: try different selector
JsonCssExtractionStrategy({
"name": "Fallback Prices",
"baseSelector": "[data-price]",
"fields": [{"name": "amount", "type": "attribute", "attribute": "data-price"}]
})
]
async with AsyncWebCrawler() as crawler:
for i, strategy in enumerate(strategies):
try:
config = CrawlerRunConfig(extraction_strategy=strategy)
result = await crawler.arun(url="https://example.com", config=config)
if result.success and result.extracted_content:
data = json.loads(result.extracted_content)
if data: # Validate non-empty results
print(f"✅ Success with strategy {i+1}: {strategy.__class__.__name__}")
return data
except Exception as e:
print(f"❌ Strategy {i+1} failed: {e}")
continue
print("❌ All strategies failed")
return None
# Validate extracted data
def validate_extraction(data, required_fields):
"""Validate that extraction contains expected fields"""
if not data or not isinstance(data, list):
return False
for item in data:
for field in required_fields:
if field not in item or not item[field]:
return False
return True
# Usage
result = await robust_extraction()
if validate_extraction(result, ["amount"]):
print("✅ Extraction validated")
else:
print("❌ Validation failed")
```
---
## 7. Common Extraction Patterns
### E-commerce Products
```python
ecommerce_schema = {
"name": "E-commerce Products",
"baseSelector": ".product, [data-product], .item",
"fields": [
{"name": "title", "selector": "h1, h2, h3, .title, .name", "type": "text"},
{"name": "price", "selector": ".price, .cost, [data-price]", "type": "text"},
{"name": "image", "selector": "img", "type": "attribute", "attribute": "src"},
{"name": "url", "selector": "a", "type": "attribute", "attribute": "href"},
{"name": "rating", "selector": ".rating, .stars", "type": "text"},
{"name": "availability", "selector": ".stock, .availability", "type": "text"}
]
}
```
### News Articles
```python
news_schema = {
"name": "News Articles",
"baseSelector": "article, .article, .post",
"fields": [
{"name": "headline", "selector": "h1, h2, .headline, .title", "type": "text"},
{"name": "author", "selector": ".author, .byline, [rel='author']", "type": "text"},
{"name": "date", "selector": "time, .date, .published", "type": "text"},
{"name": "content", "selector": ".content, .body, .text", "type": "text"},
{"name": "category", "selector": ".category, .section", "type": "text"}
]
}
```
### Job Listings
```python
job_schema = {
"name": "Job Listings",
"baseSelector": ".job, .listing, [data-job]",
"fields": [
{"name": "title", "selector": ".job-title, h2, h3", "type": "text"},
{"name": "company", "selector": ".company, .employer", "type": "text"},
{"name": "location", "selector": ".location, .place", "type": "text"},
{"name": "salary", "selector": ".salary, .pay, .compensation", "type": "text"},
{"name": "description", "selector": ".description, .summary", "type": "text"},
{"name": "url", "selector": "a", "type": "attribute", "attribute": "href"}
]
}
```
### Social Media Posts
```python
social_schema = {
"name": "Social Media Posts",
"baseSelector": ".post, .tweet, .update",
"fields": [
{"name": "username", "selector": ".username, .handle, .author", "type": "text"},
{"name": "content", "selector": ".content, .text, .message", "type": "text"},
{"name": "timestamp", "selector": ".time, .date, time", "type": "text"},
{"name": "likes", "selector": ".likes, .hearts", "type": "text"},
{"name": "shares", "selector": ".shares, .retweets", "type": "text"}
]
}
```
---
## 8. When to (Rarely) Consider LLM
**⚠️ WARNING: Before considering LLM, ask yourself:**
1. "Can I identify repeating HTML patterns?" → Use CSS/XPath schema
2. "Am I extracting simple data types?" → Use Regex patterns
3. "Can I provide a JSON example of what I want?" → Use generate_schema()
4. "Is this truly unstructured text requiring semantic understanding?" → Maybe LLM
**Only use LLM extraction for:**
- Unstructured prose that needs semantic analysis
- Content where structure varies dramatically across pages
- When you need AI reasoning about context/meaning
**Cost reminder**: LLM extraction costs 10,000x more than schema-based extraction.
---
## 9. Summary: The Extraction Hierarchy
1. **🥇 FIRST CHOICE**: `generate_schema()` - AI generates pattern once, use unlimited times
2. **🥈 SECOND CHOICE**: Manual CSS/XPath - Full control, maximum speed
3. **🥉 THIRD CHOICE**: Regex patterns - Simple data types, lightning fast
4. **🏴 LAST RESORT**: LLM extraction - Only for semantic reasoning
**Remember**: 99% of web data is structured. You almost never need LLM for extraction. Save LLM for analysis, not extraction.
**Performance**: Non-LLM strategies are 100-1000x faster and 10,000x cheaper than LLM extraction.
---
**📖 Next**: If you absolutely must use LLM extraction, see [extraction-llm.md](./extraction-llm.md) for guidance on the rare cases where it's justified.
---
## URL Seeding
Smart URL discovery for efficient large-scale crawling. Discover thousands of URLs instantly, filter by relevance, then crawl only what matters.
### Why URL Seeding vs Deep Crawling
```python
# Deep Crawling: Real-time discovery (page by page)
from crawl4ai import AsyncWebCrawler, CrawlerRunConfig
from crawl4ai.deep_crawling import BFSDeepCrawlStrategy
async def deep_crawl_example():
config = CrawlerRunConfig(
deep_crawl_strategy=BFSDeepCrawlStrategy(
max_depth=2,
include_external=False,
max_pages=50
)
)
async with AsyncWebCrawler() as crawler:
results = await crawler.arun("https://example.com", config=config)
print(f"Discovered {len(results)} pages dynamically")
# URL Seeding: Bulk discovery (thousands instantly)
from crawl4ai import AsyncUrlSeeder, SeedingConfig
async def url_seeding_example():
config = SeedingConfig(
source="sitemap+cc",
pattern="*/docs/*",
extract_head=True,
query="API documentation",
scoring_method="bm25",
max_urls=1000
)
async with AsyncUrlSeeder() as seeder:
urls = await seeder.urls("example.com", config)
print(f"Discovered {len(urls)} URLs instantly")
# Now crawl only the most relevant ones
```
### Basic URL Discovery
```python
import asyncio
from crawl4ai import AsyncUrlSeeder, SeedingConfig
async def basic_discovery():
# Context manager handles cleanup automatically
async with AsyncUrlSeeder() as seeder:
# Simple discovery from sitemaps
config = SeedingConfig(source="sitemap")
urls = await seeder.urls("example.com", config)
print(f"Found {len(urls)} URLs from sitemap")
for url in urls[:5]:
print(f" - {url['url']} (status: {url['status']})")
# Manual cleanup (if needed)
async def manual_cleanup():
seeder = AsyncUrlSeeder()
try:
config = SeedingConfig(source="cc") # Common Crawl
urls = await seeder.urls("example.com", config)
print(f"Found {len(urls)} URLs from Common Crawl")
finally:
await seeder.close()
asyncio.run(basic_discovery())
```
### Data Sources and Patterns
```python
# Different data sources
configs = [
SeedingConfig(source="sitemap"), # Fastest, official URLs
SeedingConfig(source="cc"), # Most comprehensive
SeedingConfig(source="sitemap+cc"), # Maximum coverage
]
# URL pattern filtering
patterns = [
SeedingConfig(pattern="*/blog/*"), # Blog posts only
SeedingConfig(pattern="*.html"), # HTML files only
SeedingConfig(pattern="*/product/*"), # Product pages
SeedingConfig(pattern="*/docs/api/*"), # API documentation
SeedingConfig(pattern="*"), # Everything
]
# Advanced pattern usage
async def pattern_filtering():
async with AsyncUrlSeeder() as seeder:
# Find all blog posts from 2024
config = SeedingConfig(
source="sitemap",
pattern="*/blog/2024/*.html",
max_urls=100
)
blog_urls = await seeder.urls("example.com", config)
# Further filter by keywords in URL
python_posts = [
url for url in blog_urls
if "python" in url['url'].lower()
]
print(f"Found {len(python_posts)} Python blog posts")
```
### SeedingConfig Parameters
```python
from crawl4ai import SeedingConfig
# Comprehensive configuration
config = SeedingConfig(
# Data sources
source="sitemap+cc", # "sitemap", "cc", "sitemap+cc"
pattern="*/docs/*", # URL pattern filter
# Metadata extraction
extract_head=True, # Get <head> metadata
live_check=True, # Verify URLs are accessible
# Performance controls
max_urls=1000, # Limit results (-1 = unlimited)
concurrency=20, # Parallel workers
hits_per_sec=10, # Rate limiting
# Relevance scoring
query="API documentation guide", # Search query
scoring_method="bm25", # Scoring algorithm
score_threshold=0.3, # Minimum relevance (0.0-1.0)
# Cache and filtering
force=False, # Bypass cache
filter_nonsense_urls=True, # Remove utility URLs
verbose=True # Debug output
)
# Quick configurations for common use cases
blog_config = SeedingConfig(
source="sitemap",
pattern="*/blog/*",
extract_head=True
)
api_docs_config = SeedingConfig(
source="sitemap+cc",
pattern="*/docs/*",
query="API reference documentation",
scoring_method="bm25",
score_threshold=0.5
)
product_pages_config = SeedingConfig(
source="cc",
pattern="*/product/*",
live_check=True,
max_urls=500
)
```
### Metadata Extraction and Analysis
```python
async def metadata_extraction():
async with AsyncUrlSeeder() as seeder:
config = SeedingConfig(
source="sitemap",
extract_head=True, # Extract <head> metadata
pattern="*/blog/*",
max_urls=50
)
urls = await seeder.urls("example.com", config)
# Analyze extracted metadata
for url in urls[:5]:
head_data = url['head_data']
print(f"\nURL: {url['url']}")
print(f"Title: {head_data.get('title', 'No title')}")
# Standard meta tags
meta = head_data.get('meta', {})
print(f"Description: {meta.get('description', 'N/A')}")
print(f"Keywords: {meta.get('keywords', 'N/A')}")
print(f"Author: {meta.get('author', 'N/A')}")
# Open Graph data
print(f"OG Image: {meta.get('og:image', 'N/A')}")
print(f"OG Type: {meta.get('og:type', 'N/A')}")
# JSON-LD structured data
jsonld = head_data.get('jsonld', [])
if jsonld:
print(f"Structured data: {len(jsonld)} items")
for item in jsonld[:2]:
if isinstance(item, dict):
print(f" Type: {item.get('@type', 'Unknown')}")
print(f" Name: {item.get('name', 'N/A')}")
# Filter by metadata
async def metadata_filtering():
async with AsyncUrlSeeder() as seeder:
config = SeedingConfig(
source="sitemap",
extract_head=True,
max_urls=100
)
urls = await seeder.urls("news.example.com", config)
# Filter by publication date (from JSON-LD)
from datetime import datetime, timedelta
recent_cutoff = datetime.now() - timedelta(days=7)
recent_articles = []
for url in urls:
for jsonld in url['head_data'].get('jsonld', []):
if isinstance(jsonld, dict) and 'datePublished' in jsonld:
try:
pub_date = datetime.fromisoformat(
jsonld['datePublished'].replace('Z', '+00:00')
)
if pub_date > recent_cutoff:
recent_articles.append(url)
break
except:
continue
print(f"Found {len(recent_articles)} recent articles")
```
### BM25 Relevance Scoring
```python
async def relevance_scoring():
async with AsyncUrlSeeder() as seeder:
# Find pages about Python async programming
config = SeedingConfig(
source="sitemap",
extract_head=True, # Required for content-based scoring
query="python async await concurrency",
scoring_method="bm25",
score_threshold=0.3, # Only 30%+ relevant pages
max_urls=20
)
urls = await seeder.urls("docs.python.org", config)
# Results are automatically sorted by relevance
print("Most relevant Python async content:")
for url in urls[:5]:
score = url['relevance_score']
title = url['head_data'].get('title', 'No title')
print(f"[{score:.2f}] {title}")
print(f" {url['url']}")
# URL-based scoring (when extract_head=False)
async def url_based_scoring():
async with AsyncUrlSeeder() as seeder:
config = SeedingConfig(
source="sitemap",
extract_head=False, # Fast URL-only scoring
query="machine learning tutorial",
scoring_method="bm25",
score_threshold=0.2
)
urls = await seeder.urls("example.com", config)
# Scoring based on URL structure, domain, path segments
for url in urls[:5]:
print(f"[{url['relevance_score']:.2f}] {url['url']}")
# Multi-concept queries
async def complex_queries():
queries = [
"data science pandas numpy visualization",
"web scraping automation selenium",
"machine learning tensorflow pytorch",
"api documentation rest graphql"
]
async with AsyncUrlSeeder() as seeder:
all_results = []
for query in queries:
config = SeedingConfig(
source="sitemap",
extract_head=True,
query=query,
scoring_method="bm25",
score_threshold=0.4,
max_urls=10
)
urls = await seeder.urls("learning-site.com", config)
all_results.extend(urls)
# Remove duplicates while preserving order
seen = set()
unique_results = []
for url in all_results:
if url['url'] not in seen:
seen.add(url['url'])
unique_results.append(url)
print(f"Found {len(unique_results)} unique pages across all topics")
```
### Live URL Validation
```python
async def url_validation():
async with AsyncUrlSeeder() as seeder:
config = SeedingConfig(
source="sitemap",
live_check=True, # Verify URLs are accessible
concurrency=15, # Parallel HEAD requests
hits_per_sec=8, # Rate limiting
max_urls=100
)
urls = await seeder.urls("example.com", config)
# Analyze results
valid_urls = [u for u in urls if u['status'] == 'valid']
invalid_urls = [u for u in urls if u['status'] == 'not_valid']
print(f"✅ Valid URLs: {len(valid_urls)}")
print(f"❌ Invalid URLs: {len(invalid_urls)}")
print(f"📊 Success rate: {len(valid_urls)/len(urls)*100:.1f}%")
# Show some invalid URLs for debugging
if invalid_urls:
print("\nSample invalid URLs:")
for url in invalid_urls[:3]:
print(f" - {url['url']}")
# Combined validation and metadata
async def comprehensive_validation():
async with AsyncUrlSeeder() as seeder:
config = SeedingConfig(
source="sitemap",
live_check=True, # Verify accessibility
extract_head=True, # Get metadata
query="tutorial guide", # Relevance scoring
scoring_method="bm25",
score_threshold=0.2,
concurrency=10,
max_urls=50
)
urls = await seeder.urls("docs.example.com", config)
# Filter for valid, relevant tutorials
good_tutorials = [
url for url in urls
if url['status'] == 'valid' and
url['relevance_score'] > 0.3 and
'tutorial' in url['head_data'].get('title', '').lower()
]
print(f"Found {len(good_tutorials)} high-quality tutorials")
```
### Multi-Domain Discovery
```python
async def multi_domain_research():
async with AsyncUrlSeeder() as seeder:
# Research Python tutorials across multiple sites
domains = [
"docs.python.org",
"realpython.com",
"python-course.eu",
"tutorialspoint.com"
]
config = SeedingConfig(
source="sitemap",
extract_head=True,
query="python beginner tutorial basics",
scoring_method="bm25",
score_threshold=0.3,
max_urls=15 # Per domain
)
# Discover across all domains in parallel
results = await seeder.many_urls(domains, config)
# Collect and rank all tutorials
all_tutorials = []
for domain, urls in results.items():
for url in urls:
url['domain'] = domain
all_tutorials.append(url)
# Sort by relevance across all domains
all_tutorials.sort(key=lambda x: x['relevance_score'], reverse=True)
print(f"Top 10 Python tutorials across {len(domains)} sites:")
for i, tutorial in enumerate(all_tutorials[:10], 1):
score = tutorial['relevance_score']
title = tutorial['head_data'].get('title', 'No title')[:60]
domain = tutorial['domain']
print(f"{i:2d}. [{score:.2f}] {title}")
print(f" {domain}")
# Competitor analysis
async def competitor_analysis():
competitors = ["competitor1.com", "competitor2.com", "competitor3.com"]
async with AsyncUrlSeeder() as seeder:
config = SeedingConfig(
source="sitemap",
extract_head=True,
pattern="*/blog/*",
max_urls=50
)
results = await seeder.many_urls(competitors, config)
# Analyze content strategies
for domain, urls in results.items():
content_types = {}
for url in urls:
# Extract content type from metadata
meta = url['head_data'].get('meta', {})
og_type = meta.get('og:type', 'unknown')
content_types[og_type] = content_types.get(og_type, 0) + 1
print(f"\n{domain} content distribution:")
for ctype, count in sorted(content_types.items(),
key=lambda x: x[1], reverse=True):
print(f" {ctype}: {count}")
```
### Complete Pipeline: Discovery → Filter → Crawl
```python
async def smart_research_pipeline():
"""Complete pipeline: discover URLs, filter by relevance, crawl top results"""
async with AsyncUrlSeeder() as seeder:
# Step 1: Discover relevant URLs
print("🔍 Discovering URLs...")
config = SeedingConfig(
source="sitemap+cc",
extract_head=True,
query="machine learning deep learning tutorial",
scoring_method="bm25",
score_threshold=0.4,
max_urls=100
)
urls = await seeder.urls("example.com", config)
print(f" Found {len(urls)} relevant URLs")
# Step 2: Select top articles
top_articles = sorted(urls,
key=lambda x: x['relevance_score'],
reverse=True)[:10]
print(f" Selected top {len(top_articles)} for crawling")
# Step 3: Show what we're about to crawl
print("\n📋 Articles to crawl:")
for i, article in enumerate(top_articles, 1):
score = article['relevance_score']
title = article['head_data'].get('title', 'No title')[:60]
print(f" {i}. [{score:.2f}] {title}")
# Step 4: Crawl selected articles
from crawl4ai import AsyncWebCrawler, CrawlerRunConfig
print(f"\n🕷️ Crawling {len(top_articles)} articles...")
async with AsyncWebCrawler() as crawler:
config = CrawlerRunConfig(
only_text=True,
word_count_threshold=200,
stream=True # Process results as they come
)
# Extract URLs and crawl
article_urls = [article['url'] for article in top_articles]
crawled_count = 0
async for result in await crawler.arun_many(article_urls, config=config):
if result.success:
crawled_count += 1
word_count = len(result.markdown.raw_markdown.split())
print(f" ✅ [{crawled_count}/{len(article_urls)}] "
f"{word_count} words from {result.url[:50]}...")
else:
print(f" ❌ Failed: {result.url[:50]}...")
print(f"\n✨ Successfully crawled {crawled_count} articles!")
asyncio.run(smart_research_pipeline())
```
### Advanced Features and Performance
```python
# Cache management
async def cache_management():
async with AsyncUrlSeeder() as seeder:
# First run - populate cache
config = SeedingConfig(
source="sitemap",
extract_head=True,
force=True # Bypass cache, fetch fresh
)
urls = await seeder.urls("example.com", config)
# Subsequent runs - use cache (much faster)
config = SeedingConfig(
source="sitemap",
extract_head=True,
force=False # Use cache
)
urls = await seeder.urls("example.com", config)
# Performance optimization
async def performance_tuning():
async with AsyncUrlSeeder() as seeder:
# High-performance configuration
config = SeedingConfig(
source="cc",
concurrency=50, # Many parallel workers
hits_per_sec=20, # High rate limit
max_urls=10000, # Large dataset
extract_head=False, # Skip metadata for speed
filter_nonsense_urls=True # Auto-filter utility URLs
)
import time
start = time.time()
urls = await seeder.urls("large-site.com", config)
elapsed = time.time() - start
print(f"Processed {len(urls)} URLs in {elapsed:.2f}s")
print(f"Speed: {len(urls)/elapsed:.0f} URLs/second")
# Memory-safe processing for large domains
async def large_domain_processing():
async with AsyncUrlSeeder() as seeder:
# Safe for domains with 1M+ URLs
config = SeedingConfig(
source="cc+sitemap",
concurrency=50, # Bounded queue adapts to this
max_urls=100000, # Process in batches
filter_nonsense_urls=True
)
# The seeder automatically manages memory by:
# - Using bounded queues (prevents RAM spikes)
# - Applying backpressure when queue is full
# - Processing URLs as they're discovered
urls = await seeder.urls("huge-site.com", config)
# Configuration cloning and reuse
config_base = SeedingConfig(
source="sitemap",
extract_head=True,
concurrency=20
)
# Create variations
blog_config = config_base.clone(pattern="*/blog/*")
docs_config = config_base.clone(
pattern="*/docs/*",
query="API documentation",
scoring_method="bm25"
)
fast_config = config_base.clone(
extract_head=False,
concurrency=100,
hits_per_sec=50
)
```
### Troubleshooting and Best Practices
```python
# Common issues and solutions
async def troubleshooting_guide():
async with AsyncUrlSeeder() as seeder:
# Issue: No URLs found
try:
config = SeedingConfig(source="sitemap", pattern="*/nonexistent/*")
urls = await seeder.urls("example.com", config)
if not urls:
# Solution: Try broader pattern or different source
config = SeedingConfig(source="cc+sitemap", pattern="*")
urls = await seeder.urls("example.com", config)
except Exception as e:
print(f"Discovery failed: {e}")
# Issue: Slow performance
config = SeedingConfig(
source="sitemap", # Faster than CC
concurrency=10, # Reduce if hitting rate limits
hits_per_sec=5, # Add rate limiting
extract_head=False # Skip if metadata not needed
)
# Issue: Low relevance scores
config = SeedingConfig(
query="specific detailed query terms",
score_threshold=0.1, # Lower threshold
scoring_method="bm25"
)
# Issue: Memory issues with large sites
config = SeedingConfig(
max_urls=10000, # Limit results
concurrency=20, # Reduce concurrency
source="sitemap" # Use sitemap only
)
# Performance benchmarks
print("""
Typical performance on standard connection:
- Sitemap discovery: 100-1,000 URLs/second
- Common Crawl discovery: 50-500 URLs/second
- HEAD checking: 10-50 URLs/second
- Head extraction: 5-20 URLs/second
- BM25 scoring: 10,000+ URLs/second
""")
# Best practices
best_practices = """
✅ Use context manager: async with AsyncUrlSeeder() as seeder
✅ Start with sitemaps (faster), add CC if needed
✅ Use extract_head=True only when you need metadata
✅ Set reasonable max_urls to limit processing
✅ Add rate limiting for respectful crawling
✅ Cache results with force=False for repeated operations
✅ Filter nonsense URLs (enabled by default)
✅ Use specific patterns to reduce irrelevant results
"""
```
**📖 Learn more:** [Complete URL Seeding Guide](https://docs.crawl4ai.com/core/url-seeding/), [SeedingConfig Reference](https://docs.crawl4ai.com/api/parameters/), [Multi-URL Crawling](https://docs.crawl4ai.com/advanced/multi-url-crawling/)
---
## Deep Crawling Filters & Scorers
Advanced URL filtering and scoring strategies for intelligent deep crawling with performance optimization.
### URL Filters - Content and Domain Control
```python
from crawl4ai.deep_crawling.filters import (
URLPatternFilter, DomainFilter, ContentTypeFilter,
FilterChain, ContentRelevanceFilter, SEOFilter
)
# Pattern-based filtering
pattern_filter = URLPatternFilter(
patterns=[
"*.html", # HTML pages only
"*/blog/*", # Blog posts
"*/articles/*", # Article pages
"*2024*", # Recent content
"^https://example.com/docs/.*" # Regex pattern
],
use_glob=True,
reverse=False # False = include matching, True = exclude matching
)
# Domain filtering with subdomains
domain_filter = DomainFilter(
allowed_domains=["example.com", "docs.example.com"],
blocked_domains=["ads.example.com", "tracker.com"]
)
# Content type filtering
content_filter = ContentTypeFilter(
allowed_types=["text/html", "application/pdf"],
check_extension=True
)
# Apply individual filters
url = "https://example.com/blog/2024/article.html"
print(f"Pattern filter: {pattern_filter.apply(url)}")
print(f"Domain filter: {domain_filter.apply(url)}")
print(f"Content filter: {content_filter.apply(url)}")
```
### Filter Chaining - Combine Multiple Filters
```python
# Create filter chain for comprehensive filtering
filter_chain = FilterChain([
DomainFilter(allowed_domains=["example.com"]),
URLPatternFilter(patterns=["*/blog/*", "*/docs/*"]),
ContentTypeFilter(allowed_types=["text/html"])
])
# Apply chain to URLs
urls = [
"https://example.com/blog/post1.html",
"https://spam.com/content.html",
"https://example.com/blog/image.jpg",
"https://example.com/docs/guide.html"
]
async def filter_urls(urls, filter_chain):
filtered = []
for url in urls:
if await filter_chain.apply(url):
filtered.append(url)
return filtered
# Usage
filtered_urls = await filter_urls(urls, filter_chain)
print(f"Filtered URLs: {filtered_urls}")
# Check filter statistics
for filter_obj in filter_chain.filters:
stats = filter_obj.stats
print(f"{filter_obj.name}: {stats.passed_urls}/{stats.total_urls} passed")
```
### Advanced Content Filters
```python
# BM25-based content relevance filtering
relevance_filter = ContentRelevanceFilter(
query="python machine learning tutorial",
threshold=0.5, # Minimum relevance score
k1=1.2, # TF saturation parameter
b=0.75, # Length normalization
avgdl=1000 # Average document length
)
# SEO quality filtering
seo_filter = SEOFilter(
threshold=0.65, # Minimum SEO score
keywords=["python", "tutorial", "guide"],
weights={
"title_length": 0.15,
"title_kw": 0.18,
"meta_description": 0.12,
"canonical": 0.10,
"robot_ok": 0.20,
"schema_org": 0.10,
"url_quality": 0.15
}
)
# Apply advanced filters
url = "https://example.com/python-ml-tutorial"
relevance_score = await relevance_filter.apply(url)
seo_score = await seo_filter.apply(url)
print(f"Relevance: {relevance_score}, SEO: {seo_score}")
```
### URL Scorers - Quality and Relevance Scoring
```python
from crawl4ai.deep_crawling.scorers import (
KeywordRelevanceScorer, PathDepthScorer, ContentTypeScorer,
FreshnessScorer, DomainAuthorityScorer, CompositeScorer
)
# Keyword relevance scoring
keyword_scorer = KeywordRelevanceScorer(
keywords=["python", "tutorial", "guide", "machine", "learning"],
weight=1.0,
case_sensitive=False
)
# Path depth scoring (optimal depth = 3)
depth_scorer = PathDepthScorer(
optimal_depth=3, # /category/subcategory/article
weight=0.8
)
# Content type scoring
content_type_scorer = ContentTypeScorer(
type_weights={
"html": 1.0, # Highest priority
"pdf": 0.8, # Medium priority
"txt": 0.6, # Lower priority
"doc": 0.4 # Lowest priority
},
weight=0.9
)
# Freshness scoring
freshness_scorer = FreshnessScorer(
weight=0.7,
current_year=2024
)
# Domain authority scoring
domain_scorer = DomainAuthorityScorer(
domain_weights={
"python.org": 1.0,
"github.com": 0.9,
"stackoverflow.com": 0.85,
"medium.com": 0.7,
"personal-blog.com": 0.3
},
default_weight=0.5,
weight=1.0
)
# Score individual URLs
url = "https://python.org/tutorial/2024/machine-learning.html"
scores = {
"keyword": keyword_scorer.score(url),
"depth": depth_scorer.score(url),
"content": content_type_scorer.score(url),
"freshness": freshness_scorer.score(url),
"domain": domain_scorer.score(url)
}
print(f"Individual scores: {scores}")
```
### Composite Scoring - Combine Multiple Scorers
```python
# Create composite scorer combining all strategies
composite_scorer = CompositeScorer(
scorers=[
KeywordRelevanceScorer(["python", "tutorial"], weight=1.5),
PathDepthScorer(optimal_depth=3, weight=1.0),
ContentTypeScorer({"html": 1.0, "pdf": 0.8}, weight=1.2),
FreshnessScorer(weight=0.8, current_year=2024),
DomainAuthorityScorer({
"python.org": 1.0,
"github.com": 0.9
}, weight=1.3)
],
normalize=True # Normalize by number of scorers
)
# Score multiple URLs
urls_to_score = [
"https://python.org/tutorial/2024/basics.html",
"https://github.com/user/python-guide/blob/main/README.md",
"https://random-blog.com/old/2018/python-stuff.html",
"https://python.org/docs/deep/nested/advanced/guide.html"
]
scored_urls = []
for url in urls_to_score:
score = composite_scorer.score(url)
scored_urls.append((url, score))
# Sort by score (highest first)
scored_urls.sort(key=lambda x: x[1], reverse=True)
for url, score in scored_urls:
print(f"Score: {score:.3f} - {url}")
# Check scorer statistics
print(f"\nScoring statistics:")
print(f"URLs scored: {composite_scorer.stats._urls_scored}")
print(f"Average score: {composite_scorer.stats.get_average():.3f}")
```
### Advanced Filter Patterns
```python
# Complex pattern matching
advanced_patterns = URLPatternFilter(
patterns=[
r"^https://docs\.python\.org/\d+/", # Python docs with version
r".*/tutorial/.*\.html$", # Tutorial pages
r".*/guide/(?!deprecated).*", # Guides but not deprecated
"*/blog/{2020,2021,2022,2023,2024}/*", # Recent blog posts
"**/{api,reference}/**/*.html" # API/reference docs
],
use_glob=True
)
# Exclude patterns (reverse=True)
exclude_filter = URLPatternFilter(
patterns=[
"*/admin/*",
"*/login/*",
"*/private/*",
"**/.*", # Hidden files
"*.{jpg,png,gif,css,js}$" # Media and assets
],
reverse=True # Exclude matching patterns
)
# Content type with extension mapping
detailed_content_filter = ContentTypeFilter(
allowed_types=["text", "application"],
check_extension=True,
ext_map={
"html": "text/html",
"htm": "text/html",
"md": "text/markdown",
"pdf": "application/pdf",
"doc": "application/msword",
"docx": "application/vnd.openxmlformats-officedocument.wordprocessingml.document"
}
)
```
### Performance-Optimized Filtering
```python
# High-performance filter chain for large-scale crawling
class OptimizedFilterChain:
def __init__(self):
# Fast filters first (domain, patterns)
self.fast_filters = [
DomainFilter(
allowed_domains=["example.com", "docs.example.com"],
blocked_domains=["ads.example.com"]
),
URLPatternFilter([
"*.html", "*.pdf", "*/blog/*", "*/docs/*"
])
]
# Slower filters last (content analysis)
self.slow_filters = [
ContentRelevanceFilter(
query="important content",
threshold=0.3
)
]
async def apply_optimized(self, url: str) -> bool:
# Apply fast filters first
for filter_obj in self.fast_filters:
if not filter_obj.apply(url):
return False
# Only apply slow filters if fast filters pass
for filter_obj in self.slow_filters:
if not await filter_obj.apply(url):
return False
return True
# Batch filtering with concurrency
async def batch_filter_urls(urls, filter_chain, max_concurrent=50):
import asyncio
semaphore = asyncio.Semaphore(max_concurrent)
async def filter_single(url):
async with semaphore:
return await filter_chain.apply(url), url
tasks = [filter_single(url) for url in urls]
results = await asyncio.gather(*tasks)
return [url for passed, url in results if passed]
# Usage with 1000 URLs
large_url_list = [f"https://example.com/page{i}.html" for i in range(1000)]
optimized_chain = OptimizedFilterChain()
filtered = await batch_filter_urls(large_url_list, optimized_chain)
```
### Custom Filter Implementation
```python
from crawl4ai.deep_crawling.filters import URLFilter
import re
class CustomLanguageFilter(URLFilter):
"""Filter URLs by language indicators"""
def __init__(self, allowed_languages=["en"], weight=1.0):
super().__init__()
self.allowed_languages = set(allowed_languages)
self.lang_patterns = {
"en": re.compile(r"/en/|/english/|lang=en"),
"es": re.compile(r"/es/|/spanish/|lang=es"),
"fr": re.compile(r"/fr/|/french/|lang=fr"),
"de": re.compile(r"/de/|/german/|lang=de")
}
def apply(self, url: str) -> bool:
# Default to English if no language indicators
if not any(pattern.search(url) for pattern in self.lang_patterns.values()):
result = "en" in self.allowed_languages
self._update_stats(result)
return result
# Check for allowed languages
for lang in self.allowed_languages:
if lang in self.lang_patterns:
if self.lang_patterns[lang].search(url):
self._update_stats(True)
return True
self._update_stats(False)
return False
# Custom scorer implementation
from crawl4ai.deep_crawling.scorers import URLScorer
class CustomComplexityScorer(URLScorer):
"""Score URLs by content complexity indicators"""
def __init__(self, weight=1.0):
super().__init__(weight)
self.complexity_indicators = {
"tutorial": 0.9,
"guide": 0.8,
"example": 0.7,
"reference": 0.6,
"api": 0.5
}
def _calculate_score(self, url: str) -> float:
url_lower = url.lower()
max_score = 0.0
for indicator, score in self.complexity_indicators.items():
if indicator in url_lower:
max_score = max(max_score, score)
return max_score
# Use custom filters and scorers
custom_filter = CustomLanguageFilter(allowed_languages=["en", "es"])
custom_scorer = CustomComplexityScorer(weight=1.2)
url = "https://example.com/en/tutorial/advanced-guide.html"
passes_filter = custom_filter.apply(url)
complexity_score = custom_scorer.score(url)
print(f"Passes language filter: {passes_filter}")
print(f"Complexity score: {complexity_score}")
```
### Integration with Deep Crawling
```python
from crawl4ai import AsyncWebCrawler, CrawlerRunConfig
from crawl4ai.deep_crawling import DeepCrawlStrategy
async def deep_crawl_with_filtering():
# Create comprehensive filter chain
filter_chain = FilterChain([
DomainFilter(allowed_domains=["python.org"]),
URLPatternFilter(["*/tutorial/*", "*/guide/*", "*/docs/*"]),
ContentTypeFilter(["text/html"]),
SEOFilter(threshold=0.6, keywords=["python", "programming"])
])
# Create composite scorer
scorer = CompositeScorer([
KeywordRelevanceScorer(["python", "tutorial"], weight=1.5),
FreshnessScorer(weight=0.8),
PathDepthScorer(optimal_depth=3, weight=1.0)
], normalize=True)
# Configure deep crawl strategy with filters and scorers
deep_strategy = DeepCrawlStrategy(
max_depth=3,
max_pages=100,
url_filter=filter_chain,
url_scorer=scorer,
score_threshold=0.6 # Only crawl URLs scoring above 0.6
)
config = CrawlerRunConfig(
deep_crawl_strategy=deep_strategy,
cache_mode=CacheMode.BYPASS
)
async with AsyncWebCrawler() as crawler:
result = await crawler.arun(
url="https://python.org",
config=config
)
print(f"Deep crawl completed: {result.success}")
if hasattr(result, 'deep_crawl_results'):
print(f"Pages crawled: {len(result.deep_crawl_results)}")
# Run the deep crawl
await deep_crawl_with_filtering()
```
**📖 Learn more:** [Deep Crawling Strategy](https://docs.crawl4ai.com/core/deep-crawling/), [Custom Filter Development](https://docs.crawl4ai.com/advanced/custom-filters/), [Performance Optimization](https://docs.crawl4ai.com/advanced/performance-tuning/)
---
## Summary
Crawl4AI provides a comprehensive solution for web crawling and data extraction optimized for AI applications. From simple page crawling to complex multi-URL operations with advanced filtering, the library offers the flexibility and performance needed for modern data extraction workflows.
**Key Takeaways:**
- Start with basic installation and simple crawling patterns
- Use configuration objects for consistent, maintainable code
- Choose appropriate extraction strategies based on your data structure
- Leverage Docker for production deployments
- Implement advanced features like deep crawling and custom filters as needed
**Next Steps:**
- Explore the [GitHub repository](https://github.com/unclecode/crawl4ai) for latest updates
- Join the [Discord community](https://discord.gg/jP8KfhDhyN) for support
- Check out [example projects](https://github.com/unclecode/crawl4ai/tree/main/docs/examples) for inspiration
Happy crawling! 🕷️
Reference in New Issue View Git Blame Copy Permalink