crawl4ai/docs/md_v2/assets/llm.txt/diagrams/extraction-llm.txt

## Extraction Strategy Workflows and Architecture

Visual representations of Crawl4AI's data extraction approaches, strategy selection, and processing workflows.

### Extraction Strategy Decision Tree

```mermaid
flowchart TD
    A[Content to Extract] --> B{Content Type?}
    
    B -->|Simple Patterns| C[Common Data Types]
    B -->|Structured HTML| D[Predictable Structure]
    B -->|Complex Content| E[Requires Reasoning]
    B -->|Mixed Content| F[Multiple Data Types]
    
    C --> C1{Pattern Type?}
    C1 -->|Email, Phone, URLs| C2[Built-in Regex Patterns]
    C1 -->|Custom Patterns| C3[Custom Regex Strategy]
    C1 -->|LLM-Generated| C4[One-time Pattern Generation]
    
    D --> D1{Selector Type?}
    D1 -->|CSS Selectors| D2[JsonCssExtractionStrategy]
    D1 -->|XPath Expressions| D3[JsonXPathExtractionStrategy]
    D1 -->|Need Schema?| D4[Auto-generate Schema with LLM]
    
    E --> E1{LLM Provider?}
    E1 -->|OpenAI/Anthropic| E2[Cloud LLM Strategy]
    E1 -->|Local Ollama| E3[Local LLM Strategy]
    E1 -->|Cost-sensitive| E4[Hybrid: Generate Schema Once]
    
    F --> F1[Multi-Strategy Approach]
    F1 --> F2[1. Regex for Patterns]
    F1 --> F3[2. CSS for Structure]
    F1 --> F4[3. LLM for Complex Analysis]
    
    C2 --> G[Fast Extraction ⚡]
    C3 --> G
    C4 --> H[Cached Pattern Reuse]
    
    D2 --> I[Schema-based Extraction 🏗️]
    D3 --> I
    D4 --> J[Generated Schema Cache]
    
    E2 --> K[Intelligent Parsing 🧠]
    E3 --> K
    E4 --> L[Hybrid Cost-Effective]
    
    F2 --> M[Comprehensive Results 📊]
    F3 --> M
    F4 --> M
    
    style G fill:#c8e6c9
    style I fill:#e3f2fd
    style K fill:#fff3e0
    style M fill:#f3e5f5
    style H fill:#e8f5e8
    style J fill:#e8f5e8
    style L fill:#ffecb3
```

### LLM Extraction Strategy Workflow

```mermaid
sequenceDiagram
    participant User
    participant Crawler
    participant LLMStrategy
    participant Chunker
    participant LLMProvider
    participant Parser
    
    User->>Crawler: Configure LLMExtractionStrategy
    User->>Crawler: arun(url, config)
    
    Crawler->>Crawler: Navigate to URL
    Crawler->>Crawler: Extract content (HTML/Markdown)
    Crawler->>LLMStrategy: Process content
    
    LLMStrategy->>LLMStrategy: Check content size
    
    alt Content > chunk_threshold
        LLMStrategy->>Chunker: Split into chunks with overlap
        Chunker-->>LLMStrategy: Return chunks[]
        
        loop For each chunk
            LLMStrategy->>LLMProvider: Send chunk + schema + instruction
            LLMProvider-->>LLMStrategy: Return structured JSON
        end
        
        LLMStrategy->>LLMStrategy: Merge chunk results
    else Content <= threshold
        LLMStrategy->>LLMProvider: Send full content + schema
        LLMProvider-->>LLMStrategy: Return structured JSON
    end
    
    LLMStrategy->>Parser: Validate JSON schema
    Parser-->>LLMStrategy: Validated data
    
    LLMStrategy->>LLMStrategy: Track token usage
    LLMStrategy-->>Crawler: Return extracted_content
    
    Crawler-->>User: CrawlResult with JSON data
    
    User->>LLMStrategy: show_usage()
    LLMStrategy-->>User: Token count & estimated cost
```

### Schema-Based Extraction Architecture

```mermaid
graph TB
    subgraph "Schema Definition"
        A[JSON Schema] --> A1[baseSelector]
        A --> A2[fields[]]
        A --> A3[nested structures]
        
        A2 --> A4[CSS/XPath selectors]
        A2 --> A5[Data types: text, html, attribute]
        A2 --> A6[Default values]
        
        A3 --> A7[nested objects]
        A3 --> A8[nested_list arrays]
        A3 --> A9[simple lists]
    end
    
    subgraph "Extraction Engine"
        B[HTML Content] --> C[Selector Engine]
        C --> C1[CSS Selector Parser]
        C --> C2[XPath Evaluator]
        
        C1 --> D[Element Matcher]
        C2 --> D
        
        D --> E[Type Converter]
        E --> E1[Text Extraction]
        E --> E2[HTML Preservation]
        E --> E3[Attribute Extraction]
        E --> E4[Nested Processing]
    end
    
    subgraph "Result Processing"
        F[Raw Extracted Data] --> G[Structure Builder]
        G --> G1[Object Construction]
        G --> G2[Array Assembly]
        G --> G3[Type Validation]
        
        G1 --> H[JSON Output]
        G2 --> H
        G3 --> H
    end
    
    A --> C
    E --> F
    H --> I[extracted_content]
    
    style A fill:#e3f2fd
    style C fill:#f3e5f5
    style G fill:#e8f5e8
    style H fill:#c8e6c9
```

### Automatic Schema Generation Process

```mermaid
stateDiagram-v2
    [*] --> CheckCache
    
    CheckCache --> CacheHit: Schema exists
    CheckCache --> SamplePage: Schema missing
    
    CacheHit --> LoadSchema
    LoadSchema --> FastExtraction
    
    SamplePage --> ExtractHTML: Crawl sample URL
    ExtractHTML --> LLMAnalysis: Send HTML to LLM
    LLMAnalysis --> GenerateSchema: Create CSS/XPath selectors
    GenerateSchema --> ValidateSchema: Test generated schema
    
    ValidateSchema --> SchemaWorks: Valid selectors
    ValidateSchema --> RefineSchema: Invalid selectors
    
    RefineSchema --> LLMAnalysis: Iterate with feedback
    
    SchemaWorks --> CacheSchema: Save for reuse
    CacheSchema --> FastExtraction: Use cached schema
    
    FastExtraction --> [*]: No more LLM calls needed
    
    note right of CheckCache : One-time LLM cost
    note right of FastExtraction : Unlimited fast reuse
    note right of CacheSchema : JSON file storage
```

### Multi-Strategy Extraction Pipeline

```mermaid
flowchart LR
    A[Web Page Content] --> B[Strategy Pipeline]
    
    subgraph B["Extraction Pipeline"]
        B1[Stage 1: Regex Patterns]
        B2[Stage 2: Schema-based CSS]
        B3[Stage 3: LLM Analysis]
        
        B1 --> B1a[Email addresses]
        B1 --> B1b[Phone numbers]
        B1 --> B1c[URLs and links]
        B1 --> B1d[Currency amounts]
        
        B2 --> B2a[Structured products]
        B2 --> B2b[Article metadata]
        B2 --> B2c[User reviews]
        B2 --> B2d[Navigation links]
        
        B3 --> B3a[Sentiment analysis]
        B3 --> B3b[Key topics]
        B3 --> B3c[Entity recognition]
        B3 --> B3d[Content summary]
    end
    
    B1a --> C[Result Merger]
    B1b --> C
    B1c --> C
    B1d --> C
    
    B2a --> C
    B2b --> C
    B2c --> C
    B2d --> C
    
    B3a --> C
    B3b --> C
    B3c --> C
    B3d --> C
    
    C --> D[Combined JSON Output]
    D --> E[Final CrawlResult]
    
    style B1 fill:#c8e6c9
    style B2 fill:#e3f2fd
    style B3 fill:#fff3e0
    style C fill:#f3e5f5
```

### Performance Comparison Matrix

```mermaid
graph TD
    subgraph "Strategy Performance"
        A[Extraction Strategy Comparison]
        
        subgraph "Speed ⚡"
            S1[Regex: ~10ms]
            S2[CSS Schema: ~50ms]
            S3[XPath: ~100ms]
            S4[LLM: ~2-10s]
        end
        
        subgraph "Accuracy 🎯"
            A1[Regex: Pattern-dependent]
            A2[CSS: High for structured]
            A3[XPath: Very high]
            A4[LLM: Excellent for complex]
        end
        
        subgraph "Cost 💰"
            C1[Regex: Free]
            C2[CSS: Free]
            C3[XPath: Free]
            C4[LLM: $0.001-0.01 per page]
        end
        
        subgraph "Complexity 🔧"
            X1[Regex: Simple patterns only]
            X2[CSS: Structured HTML]
            X3[XPath: Complex selectors]
            X4[LLM: Any content type]
        end
    end
    
    style S1 fill:#c8e6c9
    style S2 fill:#e8f5e8
    style S3 fill:#fff3e0
    style S4 fill:#ffcdd2
    
    style A2 fill:#e8f5e8
    style A3 fill:#c8e6c9
    style A4 fill:#c8e6c9
    
    style C1 fill:#c8e6c9
    style C2 fill:#c8e6c9
    style C3 fill:#c8e6c9
    style C4 fill:#fff3e0
    
    style X1 fill:#ffcdd2
    style X2 fill:#e8f5e8
    style X3 fill:#c8e6c9
    style X4 fill:#c8e6c9
```

### Regex Pattern Strategy Flow

```mermaid
flowchart TD
    A[Regex Extraction] --> B{Pattern Source?}
    
    B -->|Built-in| C[Use Predefined Patterns]
    B -->|Custom| D[Define Custom Regex]
    B -->|LLM-Generated| E[Generate with AI]
    
    C --> C1[Email Pattern]
    C --> C2[Phone Pattern]
    C --> C3[URL Pattern]
    C --> C4[Currency Pattern]
    C --> C5[Date Pattern]
    
    D --> D1[Write Custom Regex]
    D --> D2[Test Pattern]
    D --> D3{Pattern Works?}
    D3 -->|No| D1
    D3 -->|Yes| D4[Use Pattern]
    
    E --> E1[Provide Sample Content]
    E --> E2[LLM Analyzes Content]
    E --> E3[Generate Optimized Regex]
    E --> E4[Cache Pattern for Reuse]
    
    C1 --> F[Pattern Matching]
    C2 --> F
    C3 --> F
    C4 --> F
    C5 --> F
    D4 --> F
    E4 --> F
    
    F --> G[Extract Matches]
    G --> H[Group by Pattern Type]
    H --> I[JSON Output with Labels]
    
    style C fill:#e8f5e8
    style D fill:#e3f2fd
    style E fill:#fff3e0
    style F fill:#f3e5f5
```

### Complex Schema Structure Visualization

```mermaid
graph TB
    subgraph "E-commerce Schema Example"
        A[Category baseSelector] --> B[Category Fields]
        A --> C[Products nested_list]
        
        B --> B1[category_name]
        B --> B2[category_id attribute]
        B --> B3[category_url attribute]
        
        C --> C1[Product baseSelector]
        C1 --> C2[name text]
        C1 --> C3[price text]
        C1 --> C4[Details nested object]
        C1 --> C5[Features list]
        C1 --> C6[Reviews nested_list]
        
        C4 --> C4a[brand text]
        C4 --> C4b[model text]
        C4 --> C4c[specs html]
        
        C5 --> C5a[feature text array]
        
        C6 --> C6a[reviewer text]
        C6 --> C6b[rating attribute]
        C6 --> C6c[comment text]
        C6 --> C6d[date attribute]
    end
    
    subgraph "JSON Output Structure"
        D[categories array] --> D1[category object]
        D1 --> D2[category_name]
        D1 --> D3[category_id]
        D1 --> D4[products array]
        
        D4 --> D5[product object]
        D5 --> D6[name, price]
        D5 --> D7[details object]
        D5 --> D8[features array]
        D5 --> D9[reviews array]
        
        D7 --> D7a[brand, model, specs]
        D8 --> D8a[feature strings]
        D9 --> D9a[review objects]
    end
    
    A -.-> D
    B1 -.-> D2
    C2 -.-> D6
    C4 -.-> D7
    C5 -.-> D8
    C6 -.-> D9
    
    style A fill:#e3f2fd
    style C fill:#f3e5f5
    style C4 fill:#e8f5e8
    style D fill:#fff3e0
```

### Error Handling and Fallback Strategy

```mermaid
stateDiagram-v2
    [*] --> PrimaryStrategy
    
    PrimaryStrategy --> Success: Extraction successful
    PrimaryStrategy --> ValidationFailed: Invalid data
    PrimaryStrategy --> ExtractionFailed: No matches found
    PrimaryStrategy --> TimeoutError: LLM timeout
    
    ValidationFailed --> FallbackStrategy: Try alternative
    ExtractionFailed --> FallbackStrategy: Try alternative
    TimeoutError --> FallbackStrategy: Try alternative
    
    FallbackStrategy --> FallbackSuccess: Fallback works
    FallbackStrategy --> FallbackFailed: All strategies failed
    
    FallbackSuccess --> Success: Return results
    FallbackFailed --> ErrorReport: Log failure details
    
    Success --> [*]: Complete
    ErrorReport --> [*]: Return empty results
    
    note right of PrimaryStrategy : Try fastest/most accurate first
    note right of FallbackStrategy : Use simpler but reliable method
    note left of ErrorReport : Provide debugging information
```

### Token Usage and Cost Optimization

```mermaid
flowchart TD
    A[LLM Extraction Request] --> B{Content Size Check}
    
    B -->|Small < 1200 tokens| C[Single LLM Call]
    B -->|Large > 1200 tokens| D[Chunking Strategy]
    
    C --> C1[Send full content]
    C1 --> C2[Parse JSON response]
    C2 --> C3[Track token usage]
    
    D --> D1[Split into chunks]
    D1 --> D2[Add overlap between chunks]
    D2 --> D3[Process chunks in parallel]
    
    D3 --> D4[Chunk 1 → LLM]
    D3 --> D5[Chunk 2 → LLM]
    D3 --> D6[Chunk N → LLM]
    
    D4 --> D7[Merge results]
    D5 --> D7
    D6 --> D7
    
    D7 --> D8[Deduplicate data]
    D8 --> D9[Aggregate token usage]
    
    C3 --> E[Cost Calculation]
    D9 --> E
    
    E --> F[Usage Report]
    F --> F1[Prompt tokens: X]
    F --> F2[Completion tokens: Y]
    F --> F3[Total cost: $Z]
    
    style C fill:#c8e6c9
    style D fill:#fff3e0
    style E fill:#e3f2fd
    style F fill:#f3e5f5
```

**📖 Learn more:** [LLM Strategies](https://docs.crawl4ai.com/extraction/llm-strategies/), [Schema-Based Extraction](https://docs.crawl4ai.com/extraction/no-llm-strategies/), [Pattern Matching](https://docs.crawl4ai.com/extraction/no-llm-strategies/#regexextractionstrategy), [Performance Optimization](https://docs.crawl4ai.com/advanced/multi-url-crawling/)