Feat: advanced markdown parsing (#9607)

### What problem does this PR solve?

Using AST parsing to handle markdown more accurately, preventing
components from being cut off by chunking. #9564

<img width="1746" height="993" alt="image"
src="https://github.com/user-attachments/assets/4aaf4bf6-5714-4d48-a9cf-864f59633f7f"
/>

<img width="1739" height="982" alt="image"
src="https://github.com/user-attachments/assets/dc00233f-7a55-434f-bbb7-74ce7f57a6cf"
/>

<img width="559" height="100" alt="image"
src="https://github.com/user-attachments/assets/4a556b5b-d9c6-4544-a486-8ac342bd504e"
/>


### Type of change

- [x] New Feature (non-breaking change which adds functionality)

deepdoc/parser/__init__.py

@@ -14,13 +14,15 @@
 #  limitations under the License.
 #
 
-from .pdf_parser import RAGFlowPdfParser as PdfParser, PlainParser
 from .docx_parser import RAGFlowDocxParser as DocxParser
 from .excel_parser import RAGFlowExcelParser as ExcelParser
-from .ppt_parser import RAGFlowPptParser as PptParser
 from .html_parser import RAGFlowHtmlParser as HtmlParser
 from .json_parser import RAGFlowJsonParser as JsonParser
+from .markdown_parser import MarkdownElementExtractor
 from .markdown_parser import RAGFlowMarkdownParser as MarkdownParser
+from .pdf_parser import PlainParser
+from .pdf_parser import RAGFlowPdfParser as PdfParser
+from .ppt_parser import RAGFlowPptParser as PptParser
 from .txt_parser import RAGFlowTxtParser as TxtParser
 
 __all__ = [
@@ -33,4 +35,6 @@ __all__ = [
     "JsonParser",
     "MarkdownParser",
     "TxtParser",
+    "MarkdownElementExtractor",
 ]
+

deepdoc/parser/markdown_parser.py

@@ -17,8 +17,10 @@
 
 import re
 
+import mistune
 from markdown import markdown
 
+
 class RAGFlowMarkdownParser:
     def __init__(self, chunk_token_num=128):
         self.chunk_token_num = int(chunk_token_num)
@@ -38,7 +40,7 @@ class RAGFlowMarkdownParser:
                     new_text += working_text[last_end : match.start()] + "\n\n"
                 else:
                     # Replace with rendered HTML
-                    html_table = markdown(raw_table, extensions=['markdown.extensions.tables']) if render else raw_table
+                    html_table = markdown(raw_table, extensions=["markdown.extensions.tables"]) if render else raw_table
                     new_text += working_text[last_end : match.start()] + html_table + "\n\n"
                 last_end = match.end()
             new_text += working_text[last_end:]
@@ -47,28 +49,32 @@ class RAGFlowMarkdownParser:
         if "|" in markdown_text:  # for optimize performance
             # Standard Markdown table
             border_table_pattern = re.compile(
-                r'''
+                r"""
                 (?:\n|^)
                 (?:\|.*?\|.*?\|.*?\n)
                 (?:\|(?:\s*[:-]+[-| :]*\s*)\|.*?\n)
                 (?:\|.*?\|.*?\|.*?\n)+
-            ''', re.VERBOSE)
+            """,
+                re.VERBOSE,
+            )
             working_text = replace_tables_with_rendered_html(border_table_pattern, tables)
 
             # Borderless Markdown table
             no_border_table_pattern = re.compile(
-                r'''
+                r"""
                 (?:\n|^)
                 (?:\S.*?\|.*?\n)
                 (?:(?:\s*[:-]+[-| :]*\s*).*?\n)
                 (?:\S.*?\|.*?\n)+
-                ''', re.VERBOSE)
+                """,
+                re.VERBOSE,
+            )
             working_text = replace_tables_with_rendered_html(no_border_table_pattern, tables)
 
         if "<table>" in working_text.lower():  # for optimize performance
             # HTML table extraction - handle possible html/body wrapper tags
             html_table_pattern = re.compile(
-            r'''
+                r"""
             (?:\n|^)
             \s*
             (?:
@@ -83,9 +89,10 @@ class RAGFlowMarkdownParser:
             )
             \s*
             (?=\n|$)
-            ''',
+            """,
-            re.VERBOSE | re.DOTALL | re.IGNORECASE
+                re.VERBOSE | re.DOTALL | re.IGNORECASE,
             )
+
             def replace_html_tables():
                 nonlocal working_text
                 new_text = ""
@@ -104,3 +111,163 @@ class RAGFlowMarkdownParser:
             replace_html_tables()
 
         return working_text, tables
+
+
+class MarkdownElementExtractor:
+    def __init__(self, markdown_content):
+        self.markdown_content = markdown_content
+        self.lines = markdown_content.split("\n")
+        self.ast_parser = mistune.create_markdown(renderer="ast")
+        self.ast_nodes = self.ast_parser(markdown_content)
+
+    def extract_elements(self):
+        """Extract individual elements (headers, code blocks, lists, etc.)"""
+        sections = []
+
+        i = 0
+        while i < len(self.lines):
+            line = self.lines[i]
+
+            if re.match(r"^#{1,6}\s+.*$", line):
+                # header
+                element = self._extract_header(i)
+                sections.append(element["content"])
+                i = element["end_line"] + 1
+            elif line.strip().startswith("```"):
+                # code block
+                element = self._extract_code_block(i)
+                sections.append(element["content"])
+                i = element["end_line"] + 1
+            elif re.match(r"^\s*[-*+]\s+.*$", line) or re.match(r"^\s*\d+\.\s+.*$", line):
+                # list block
+                element = self._extract_list_block(i)
+                sections.append(element["content"])
+                i = element["end_line"] + 1
+            elif line.strip().startswith(">"):
+                # blockquote
+                element = self._extract_blockquote(i)
+                sections.append(element["content"])
+                i = element["end_line"] + 1
+            elif line.strip():
+                # text block (paragraphs and inline elements until next block element)
+                element = self._extract_text_block(i)
+                sections.append(element["content"])
+                i = element["end_line"] + 1
+            else:
+                i += 1
+
+        sections = [section for section in sections if section.strip()]
+        return sections
+
+    def _extract_header(self, start_pos):
+        return {
+            "type": "header",
+            "content": self.lines[start_pos],
+            "start_line": start_pos,
+            "end_line": start_pos,
+        }
+
+    def _extract_code_block(self, start_pos):
+        end_pos = start_pos
+        content_lines = [self.lines[start_pos]]
+
+        # Find the end of the code block
+        for i in range(start_pos + 1, len(self.lines)):
+            content_lines.append(self.lines[i])
+            end_pos = i
+            if self.lines[i].strip().startswith("```"):
+                break
+
+        return {
+            "type": "code_block",
+            "content": "\n".join(content_lines),
+            "start_line": start_pos,
+            "end_line": end_pos,
+        }
+
+    def _extract_list_block(self, start_pos):
+        end_pos = start_pos
+        content_lines = []
+
+        i = start_pos
+        while i < len(self.lines):
+            line = self.lines[i]
+            # check if this line is a list item or continuation of a list
+            if (
+                re.match(r"^\s*[-*+]\s+.*$", line)
+                or re.match(r"^\s*\d+\.\s+.*$", line)
+                or (i > start_pos and not line.strip())
+                or (i > start_pos and re.match(r"^\s{2,}[-*+]\s+.*$", line))
+                or (i > start_pos and re.match(r"^\s{2,}\d+\.\s+.*$", line))
+                or (i > start_pos and re.match(r"^\s+\w+.*$", line))
+            ):
+                content_lines.append(line)
+                end_pos = i
+                i += 1
+            else:
+                break
+
+        return {
+            "type": "list_block",
+            "content": "\n".join(content_lines),
+            "start_line": start_pos,
+            "end_line": end_pos,
+        }
+
+    def _extract_blockquote(self, start_pos):
+        end_pos = start_pos
+        content_lines = []
+
+        i = start_pos
+        while i < len(self.lines):
+            line = self.lines[i]
+            if line.strip().startswith(">") or (i > start_pos and not line.strip()):
+                content_lines.append(line)
+                end_pos = i
+                i += 1
+            else:
+                break
+
+        return {
+            "type": "blockquote",
+            "content": "\n".join(content_lines),
+            "start_line": start_pos,
+            "end_line": end_pos,
+        }
+
+    def _extract_text_block(self, start_pos):
+        """Extract a text block (paragraphs, inline elements) until next block element"""
+        end_pos = start_pos
+        content_lines = [self.lines[start_pos]]
+
+        i = start_pos + 1
+        while i < len(self.lines):
+            line = self.lines[i]
+            # stop if we encounter a block element
+            if re.match(r"^#{1,6}\s+.*$", line) or line.strip().startswith("```") or re.match(r"^\s*[-*+]\s+.*$", line) or re.match(r"^\s*\d+\.\s+.*$", line) or line.strip().startswith(">"):
+                break
+            elif not line.strip():
+                # check if the next line is a block element
+                if i + 1 < len(self.lines) and (
+                    re.match(r"^#{1,6}\s+.*$", self.lines[i + 1])
+                    or self.lines[i + 1].strip().startswith("```")
+                    or re.match(r"^\s*[-*+]\s+.*$", self.lines[i + 1])
+                    or re.match(r"^\s*\d+\.\s+.*$", self.lines[i + 1])
+                    or self.lines[i + 1].strip().startswith(">")
+                ):
+                    break
+                else:
+                    content_lines.append(line)
+                    end_pos = i
+                    i += 1
+            else:
+                content_lines.append(line)
+                end_pos = i
+                i += 1
+
+        return {
+            "type": "text_block",
+            "content": "\n".join(content_lines),
+            "start_line": start_pos,
+            "end_line": end_pos,
+        }

rag/app/naive.py

@@ -30,7 +30,7 @@ from tika import parser
 
 from api.db import LLMType
 from api.db.services.llm_service import LLMBundle
-from deepdoc.parser import DocxParser, ExcelParser, HtmlParser, JsonParser, MarkdownParser, PdfParser, TxtParser
+from deepdoc.parser import DocxParser, ExcelParser, HtmlParser, JsonParser, MarkdownElementExtractor, MarkdownParser, PdfParser, TxtParser
 from deepdoc.parser.figure_parser import VisionFigureParser, vision_figure_parser_figure_data_wrapper
 from deepdoc.parser.pdf_parser import PlainParser, VisionParser
 from rag.nlp import concat_img, find_codec, naive_merge, naive_merge_with_images, naive_merge_docx, rag_tokenizer, tokenize_chunks, tokenize_chunks_with_images, tokenize_table
@@ -350,17 +350,14 @@ class Markdown(MarkdownParser):
         else:
             with open(filename, "r") as f:
                 txt = f.read()
+
         remainder, tables = self.extract_tables_and_remainder(f'{txt}\n', separate_tables=separate_tables)
-        sections = []
+
+        extractor = MarkdownElementExtractor(txt)
+        element_sections = extractor.extract_elements()
+        sections = [(element, "") for element in element_sections]
+
         tbls = []
-        for sec in remainder.split("\n"):
-            if sec.strip().find("#") == 0:
-                sections.append((sec, ""))
-            elif sections and sections[-1][0].strip().find("#") == 0:
-                sec_, _ = sections.pop(-1)
-                sections.append((sec_ + "\n" + sec, ""))
-            else:
-                sections.append((sec, ""))
         for table in tables:
             tbls.append(((None, markdown(table, extensions=['markdown.extensions.tables'])), ""))
         return sections, tbls