fix: Add dedicated document LLM, improve Markdown detection and conte…

src/memos/api/config.py

@@ -400,7 +400,7 @@ def get_activation_config() -> dict[str, Any]:
 
     @staticmethod
     def get_memreader_config() -> dict[str, Any]:
-        """Get MemReader configuration for chat/doc extraction (fine-tuned 0.6B model).
+        """Get the main MemReader configuration for text/chat extraction.
 
         When MEMREADER_GENERAL_MODEL is configured (i.e. a separate stable LLM exists),
         the backup client is automatically enabled so that primary failures (self-deployed
@@ -444,7 +444,7 @@ def get_qwen_llm_config() -> dict[str, Any] | None:
 
     @staticmethod
     def get_memreader_general_llm_config() -> dict[str, Any]:
-        """Get general LLM configuration for non-chat/doc tasks.
+        """Get general LLM configuration for non-primary extraction tasks.
 
         Used for: hallucination filter, memory rewrite, memory merge,
         tool trajectory extraction, skill memory extraction.
@@ -478,6 +478,24 @@ def get_image_parser_llm_config() -> dict[str, Any]:
         # Fallback to general_llm config (which itself falls back to OpenAI)
         return APIConfig.get_memreader_general_llm_config()
 
+    @staticmethod
+    def get_document_parser_llm_config() -> dict[str, Any] | None:
+        """Get the dedicated LLM configuration for document extraction.
+
+        The provider endpoint and credentials are selected from QWEN_* or
+        OPENAI_* according to DOCUMENT_PARSER_MODEL.
+
+        Fallback chain: DOCUMENT_PARSER_MODEL -> MEMREADER_GENERAL_MODEL.
+        Returns None when neither model is configured.
+        """
+        document_model = os.getenv("DOCUMENT_PARSER_MODEL") or os.getenv("MEMREADER_GENERAL_MODEL")
+        if not document_model:
+            return None
+        return APIConfig._build_provider_llm_config(
+            document_model,
+            temperature=0.8,
+        )
+
     @staticmethod
     def get_preference_extractor_llm_config() -> dict[str, Any]:
         """Get LLM configuration for preference extraction.
@@ -994,6 +1012,8 @@ def get_product_default_config() -> dict[str, Any]:
                     "general_llm": APIConfig.get_memreader_general_llm_config(),
                     # Image parser LLM (requires vision model)
                     "image_parser_llm": APIConfig.get_image_parser_llm_config(),
+                    # Dedicated LLM for document chunk extraction
+                    "document_parser_llm": APIConfig.get_document_parser_llm_config(),
                     # Preference extractor LLM. Reader falls back to general_llm when unset.
                     "preference_extractor_llm": APIConfig.get_preference_extractor_llm_config()
                     if os.getenv("PREFERENCE_EXTRACTOR_MODEL")
@@ -1127,6 +1147,8 @@ def create_user_config(user_name: str, user_id: str) -> tuple["MOSConfig", "Gene
                     "general_llm": APIConfig.get_memreader_general_llm_config(),
                     # Image parser LLM (requires vision model)
                     "image_parser_llm": APIConfig.get_image_parser_llm_config(),
+                    # Dedicated LLM for document chunk extraction
+                    "document_parser_llm": APIConfig.get_document_parser_llm_config(),
                     # Preference extractor LLM. Reader falls back to general_llm when unset.
                     "preference_extractor_llm": APIConfig.get_preference_extractor_llm_config()
                     if os.getenv("PREFERENCE_EXTRACTOR_MODEL")

src/memos/configs/mem_reader.py

@@ -25,7 +25,8 @@ def parse_datetime(cls, value):
         return value
 
     llm: LLMConfigFactory = Field(
-        ..., description="LLM configuration for chat/doc memory extraction (fine-tuned model)"
+        ...,
+        description="Main LLM configuration for standard text/chat memory extraction",
     )
     general_llm: LLMConfigFactory | None = Field(
         default=None,
@@ -36,6 +37,10 @@ def parse_datetime(cls, value):
         default=None,
         description="Vision LLM for image parsing. Falls back to general_llm if not set.",
     )
+    document_parser_llm: LLMConfigFactory | None = Field(
+        default=None,
+        description="Dedicated LLM for document content extraction",
+    )
     preference_extractor_llm: LLMConfigFactory | None = Field(
         default=None,
         description="LLM for preference extraction. Falls back to general_llm if not set.",

src/memos/mem_reader/multi_modal_struct.py

@@ -71,12 +71,19 @@ def __init__(self, config: MultiModalStructMemReaderConfig):
             if config.image_parser_llm is not None
             else self.general_llm
         )
+        # Document parser LLM (separate from the fine-tuned chat extractor)
+        self.document_parser_llm = (
+            LLMFactory.from_config(config.document_parser_llm)
+            if config.document_parser_llm is not None
+            else None
+        )
         # Initialize MultiModalParser for routing to different parsers
-        # Pass image_parser_llm for image parsing
+        # Pass dedicated image/document LLMs to their corresponding parsers
         self.multi_modal_parser = MultiModalParser(
             embedder=self.embedder,
             llm=self.llm,
             image_parser_llm=self.image_parser_llm,
+            document_parser_llm=self.document_parser_llm,
             parser=None,
             direct_markdown_hostnames=direct_markdown_hostnames,
         )

src/memos/mem_reader/read_multi_modal/file_content_parser.py

@@ -100,7 +100,7 @@ def _get_doc_llm_response(
     def _handle_url(self, url_str: str, filename: str) -> tuple[str, str | None, bool]:
         """Download and parse file from URL."""
         try:
-            from urllib.parse import urlparse
+            from urllib.parse import unquote, urlparse
 
             import requests
 
@@ -118,7 +118,13 @@ def _handle_url(self, url_str: str, filename: str) -> tuple[str, str | None, boo
                 return response.text, None, True
 
             file_ext = os.path.splitext(filename)[1].lower()
-            if file_ext in [".md", ".markdown", ".txt"] or self._is_oss_md(url_str):
+            url_path_ext = os.path.splitext(unquote(parsed_url.path))[1].lower()
+            markdown_extensions = {".md", ".markdown", ".txt"}
+            if (
+                file_ext in markdown_extensions
+                or url_path_ext in markdown_extensions
+                or self._is_oss_md(url_str)
+            ):
                 return response.text, None, True
             with tempfile.NamedTemporaryFile(mode="wb", delete=False, suffix=file_ext) as temp_file:
                 temp_file.write(response.content)
@@ -333,7 +339,7 @@ def __init__(
 
         Args:
             embedder: Embedder for generating embeddings
-            llm: Optional LLM for fine mode processing
+            llm: Optional dedicated LLM for document extraction
             parser: Optional parser for parsing file contents
             direct_markdown_hostnames: List of hostnames that should return markdown directly
                 without parsing. If None, reads from FILE_PARSER_DIRECT_MARKDOWN_HOSTNAMES
@@ -739,6 +745,13 @@ def parse_fine(
             logger.info(
                 f"[Chunker: FileContentParser] Extracted {len(headers)} headers from markdown"
             )
+            document_context = self._build_markdown_document_context(headers, filename)
+            if document_context:
+                context_parts = []
+                if message_text_context:
+                    context_parts.append(f"Related message context:\n{message_text_context}")
+                context_parts.append(document_context)
+                message_text_context = "\n\n".join(context_parts)
 
         # Extract and process images from parsed_text
         if is_markdown and parsed_text and self.image_parser:
@@ -1050,6 +1063,37 @@ def _extract_markdown_headers(self, text: str) -> dict[int, dict]:
         logger.info(f"[Chunker: FileContentParser] Extracted {len(headers)} headers from markdown")
         return headers
 
+    @staticmethod
+    def _build_markdown_document_context(
+        headers: dict[int, dict],
+        filename: str,
+    ) -> str | None:
+        """Build document-level context from a Markdown filename and H1 titles."""
+        h1_titles = []
+        seen_titles = set()
+        for header in headers.values():
+            if header.get("level") != 1:
+                continue
+            title = str(header.get("title", "")).strip()
+            if not title or title in seen_titles:
+                continue
+            seen_titles.add(title)
+            h1_titles.append(title)
+
+        if not filename and not h1_titles:
+            return None
+
+        lines = [
+            "Document-level context for disambiguation only. "
+            "Do not create memories from this context unless the current chunk supports them."
+        ]
+        if filename:
+            lines.append(f"Filename: {filename}")
+        if h1_titles:
+            lines.append("Top-level Markdown headings:")
+            lines.extend(f"- {title}" for title in h1_titles)
+        return "\n".join(lines)
+
     def _get_header_context(
         self, text: str, image_position: int, headers: dict[int, dict]
     ) -> list[str]:

src/memos/mem_reader/read_multi_modal/multi_modal_parser.py

@@ -38,6 +38,7 @@ def __init__(
         embedder: BaseEmbedder,
         llm: BaseLLM | None = None,
         image_parser_llm: BaseLLM | None = None,
+        document_parser_llm: BaseLLM | None = None,
         parser: Any | None = None,
         direct_markdown_hostnames: list[str] | None = None,
     ):
@@ -46,9 +47,10 @@ def __init__(
 
         Args:
             embedder: Embedder for generating embeddings
-            llm: Optional LLM for fine mode processing (chat/doc extraction)
+            llm: Optional main LLM for standard text/chat extraction.
             image_parser_llm: Optional vision LLM for image parsing.
                 Falls back to llm if not provided.
+            document_parser_llm: Optional dedicated LLM for document extraction.
             parser: Optional parser for parsing file contents
             direct_markdown_hostnames: List of hostnames that should return markdown directly
                 without parsing. If None, reads from FILE_PARSER_DIRECT_MARKDOWN_HOSTNAMES
@@ -58,6 +60,7 @@ def __init__(
         self.llm = llm
         # Image parser LLM (requires vision model), falls back to main llm
         self.image_parser_llm = image_parser_llm if image_parser_llm is not None else llm
+        self.document_parser_llm = document_parser_llm
         self.parser = parser
 
         # Initialize parsers for different message types
@@ -71,7 +74,7 @@ def __init__(
         self.image_parser = ImageParser(embedder, self.image_parser_llm)
         self.file_content_parser = FileContentParser(
             embedder,
-            llm,
+            self.document_parser_llm,
             parser,
             direct_markdown_hostnames=direct_markdown_hostnames,
             image_parser=self.image_parser,