In this commit, we introduce the new concept of MakrdownGenerationStrategy, which allows us to expand our future strategies to generate better markdown. Right now, we generate raw markdown as we were doing before. We have a new algorithm for fitting markdown based on BM25, and now we add the ability to refine markdown into a citation form. Our links will be extracted and replaced by a citation reference number, and then we will have reference sections at the very end; we add all the links with the descriptions. This format is more suitable for large language models. In case we don't need to pass links, we can reduce the size of the markdown significantly and also attach the list of references as a separate file to a large language model. This commit contains changes for this direction.

crawl4ai/models.py

@@ -1,5 +1,5 @@
 from pydantic import BaseModel, HttpUrl
-from typing import List, Dict, Optional, Callable, Awaitable
+from typing import List, Dict, Optional, Callable, Awaitable, Union
 
 
 
@@ -7,6 +7,12 @@ class UrlModel(BaseModel):
     url: HttpUrl
     forced: bool = False
 
+class MarkdownGenerationResult(BaseModel):
+    raw_markdown: str
+    markdown_with_citations: str
+    references_markdown: str
+    fit_markdown: Optional[str] = None
+
 class CrawlResult(BaseModel):
     url: str
     html: str
@@ -16,7 +22,8 @@ class CrawlResult(BaseModel):
     links: Dict[str, List[Dict]] = {}
     downloaded_files: Optional[List[str]] = None
     screenshot: Optional[str] = None
-    markdown: Optional[str] = None
+    markdown: Optional[Union[str, MarkdownGenerationResult]] = None
+    markdown_v2: Optional[MarkdownGenerationResult] = None
     fit_markdown: Optional[str] = None
     fit_html: Optional[str] = None
     extracted_content: Optional[str] = None
@@ -36,3 +43,5 @@ class AsyncCrawlResponse(BaseModel):
 
     class Config:
         arbitrary_types_allowed = True
+
+