Add Chain of Thought support to Gemini LLM integration

- Extract thoughts from response parts
- Add COT enable/disable parameter

lightrag/llm/gemini.py

@@ -114,28 +114,44 @@ def _format_history_messages(history_messages: list[dict[str, Any]] | None) -> s
     return "\n".join(history_lines)
 
 
-def _extract_response_text(response: Any) -> str:
+def _extract_response_text(
+    response: Any, extract_thoughts: bool = False
+) -> tuple[str, str]:
     """
-    Extract text content from Gemini response, avoiding warnings about non-text parts.
+    Extract text content from Gemini response, separating regular content from thoughts.
 
-    Always extracts text manually from parts to avoid triggering warnings when
-    non-text parts (like 'thought_signature') are present in the response.
+    Args:
+        response: Gemini API response object
+        extract_thoughts: Whether to extract thought content separately
+
+    Returns:
+        Tuple of (regular_text, thought_text)
     """
     candidates = getattr(response, "candidates", None)
     if not candidates:
-        return ""
+        return ("", "")
+
+    regular_parts: list[str] = []
+    thought_parts: list[str] = []
 
-    parts: list[str] = []
     for candidate in candidates:
         if not getattr(candidate, "content", None):
             continue
-        for part in getattr(candidate.content, "parts", []):
-            # Only extract text parts to avoid non-text content like thought_signature
+        # Use 'or []' to handle None values from parts attribute
+        for part in getattr(candidate.content, "parts", None) or []:
             text = getattr(part, "text", None)
-            if text:
-                parts.append(text)
+            if not text:
+                continue
 
-    return "\n".join(parts)
+            # Check if this part is thought content using the 'thought' attribute
+            is_thought = getattr(part, "thought", False)
+
+            if is_thought and extract_thoughts:
+                thought_parts.append(text)
+            elif not is_thought:
+                regular_parts.append(text)
+
+    return ("\n".join(regular_parts), "\n".join(thought_parts))
 
 
 async def gemini_complete_if_cache(
@@ -143,18 +159,51 @@ async def gemini_complete_if_cache(
     prompt: str,
     system_prompt: str | None = None,
     history_messages: list[dict[str, Any]] | None = None,
-    *,
-    api_key: str | None = None,
+    enable_cot: bool = False,
     base_url: str | None = None,
-    generation_config: dict[str, Any] | None = None,
-    keyword_extraction: bool = False,
+    api_key: str | None = None,
     token_tracker: Any | None = None,
-    hashing_kv: Any | None = None,  # noqa: ARG001 - present for interface parity
     stream: bool | None = None,
-    enable_cot: bool = False,  # noqa: ARG001 - not supported by Gemini currently
-    timeout: float | None = None,  # noqa: ARG001 - handled by caller if needed
+    keyword_extraction: bool = False,
+    generation_config: dict[str, Any] | None = None,
     **_: Any,
 ) -> str | AsyncIterator[str]:
+    """
+    Complete a prompt using Gemini's API with Chain of Thought (COT) support.
+
+    This function supports automatic integration of reasoning content from Gemini models
+    that provide Chain of Thought capabilities via the thinking_config API feature.
+
+    COT Integration:
+    - When enable_cot=True: Thought content is wrapped in <think>...</think> tags
+    - When enable_cot=False: Thought content is filtered out, only regular content returned
+    - Thought content is identified by the 'thought' attribute on response parts
+    - Requires thinking_config to be enabled in generation_config for API to return thoughts
+
+    Args:
+        model: The Gemini model to use.
+        prompt: The prompt to complete.
+        system_prompt: Optional system prompt to include.
+        history_messages: Optional list of previous messages in the conversation.
+        api_key: Optional Gemini API key. If None, uses environment variable.
+        base_url: Optional custom API endpoint.
+        generation_config: Optional generation configuration dict.
+        keyword_extraction: Whether to use JSON response format.
+        token_tracker: Optional token usage tracker for monitoring API usage.
+        hashing_kv: Storage interface (for interface parity with other bindings).
+        stream: Whether to stream the response.
+        enable_cot: Whether to include Chain of Thought content in the response.
+        timeout: Request timeout (handled by caller if needed).
+        **_: Additional keyword arguments (ignored).
+
+    Returns:
+        The completed text (with COT content if enable_cot=True) or an async iterator
+        of text chunks if streaming. COT content is wrapped in <think>...</think> tags.
+
+    Raises:
+        RuntimeError: If the response from Gemini is empty.
+        ValueError: If API key is not provided or configured.
+    """
     loop = asyncio.get_running_loop()
 
     key = _ensure_api_key(api_key)
@@ -188,6 +237,11 @@ async def gemini_complete_if_cache(
         usage_container: dict[str, Any] = {}
 
         def _stream_model() -> None:
+            # COT state tracking for streaming
+            cot_active = False
+            cot_started = False
+            initial_content_seen = False
+
             try:
                 stream_kwargs = dict(request_kwargs)
                 stream_iterator = client.models.generate_content_stream(**stream_kwargs)
@@ -195,19 +249,59 @@ async def gemini_complete_if_cache(
                     usage = getattr(chunk, "usage_metadata", None)
                     if usage is not None:
                         usage_container["usage"] = usage
-                    # Always use manual extraction to avoid warnings about non-text parts
-                    text_piece = _extract_response_text(chunk)
-                    if text_piece:
-                        loop.call_soon_threadsafe(queue.put_nowait, text_piece)
+
+                    # Extract both regular and thought content
+                    regular_text, thought_text = _extract_response_text(
+                        chunk, extract_thoughts=True
+                    )
+
+                    if enable_cot:
+                        # Process regular content
+                        if regular_text:
+                            if not initial_content_seen:
+                                initial_content_seen = True
+
+                            # Close COT section if it was active
+                            if cot_active:
+                                loop.call_soon_threadsafe(queue.put_nowait, "</think>")
+                                cot_active = False
+
+                            # Send regular content
+                            loop.call_soon_threadsafe(queue.put_nowait, regular_text)
+
+                        # Process thought content
+                        if thought_text:
+                            if not initial_content_seen and not cot_started:
+                                # Start COT section
+                                loop.call_soon_threadsafe(queue.put_nowait, "<think>")
+                                cot_active = True
+                                cot_started = True
+
+                            # Send thought content if COT is active
+                            if cot_active:
+                                loop.call_soon_threadsafe(queue.put_nowait, thought_text)
+                    else:
+                        # COT disabled - only send regular content
+                        if regular_text:
+                            loop.call_soon_threadsafe(queue.put_nowait, regular_text)
+
+                # Ensure COT is properly closed if still active
+                if cot_active:
+                    loop.call_soon_threadsafe(queue.put_nowait, "</think>")
+
                 loop.call_soon_threadsafe(queue.put_nowait, None)
             except Exception as exc:  # pragma: no cover - surface runtime issues
+                # Try to close COT tag before reporting error
+                if cot_active:
+                    try:
+                        loop.call_soon_threadsafe(queue.put_nowait, "</think>")
+                    except Exception:
+                        pass
                 loop.call_soon_threadsafe(queue.put_nowait, exc)
 
         loop.run_in_executor(None, _stream_model)
 
         async def _async_stream() -> AsyncIterator[str]:
-            accumulated = ""
-            emitted = ""
             try:
                 while True:
                     item = await queue.get()
@@ -220,16 +314,9 @@ async def gemini_complete_if_cache(
                     if "\\u" in chunk_text:
                         chunk_text = safe_unicode_decode(chunk_text.encode("utf-8"))
 
-                    accumulated += chunk_text
-                    sanitized = remove_think_tags(accumulated)
-                    if sanitized.startswith(emitted):
-                        delta = sanitized[len(emitted) :]
-                    else:
-                        delta = sanitized
-                    emitted = sanitized
-
-                    if delta:
-                        yield delta
+                    # Yield the chunk directly without filtering
+                    # COT filtering is already handled in _stream_model()
+                    yield chunk_text
             finally:
                 usage = usage_container.get("usage")
                 if token_tracker and usage:
@@ -247,14 +334,33 @@ async def gemini_complete_if_cache(
 
     response = await asyncio.to_thread(_call_model)
 
-    text = _extract_response_text(response)
-    if not text:
+    # Extract both regular text and thought text
+    regular_text, thought_text = _extract_response_text(response, extract_thoughts=True)
+
+    # Apply COT filtering logic based on enable_cot parameter
+    if enable_cot:
+        # Include thought content wrapped in <think> tags
+        if thought_text and thought_text.strip():
+            if not regular_text or regular_text.strip() == "":
+                # Only thought content available
+                final_text = f"<think>{thought_text}</think>"
+            else:
+                # Both content types present: prepend thought to regular content
+                final_text = f"<think>{thought_text}</think>{regular_text}"
+        else:
+            # No thought content, use regular content only
+            final_text = regular_text or ""
+    else:
+        # Filter out thought content, return only regular content
+        final_text = regular_text or ""
+
+    if not final_text:
         raise RuntimeError("Gemini response did not contain any text content.")
 
-    if "\\u" in text:
-        text = safe_unicode_decode(text.encode("utf-8"))
+    if "\\u" in final_text:
+        final_text = safe_unicode_decode(final_text.encode("utf-8"))
 
-    text = remove_think_tags(text)
+    final_text = remove_think_tags(final_text)
 
     usage = getattr(response, "usage_metadata", None)
     if token_tracker and usage:
@@ -266,8 +372,8 @@ async def gemini_complete_if_cache(
             }
         )
 
-    logger.debug("Gemini response length: %s", len(text))
-    return text
+    logger.debug("Gemini response length: %s", len(final_text))
+    return final_text
 
 
 async def gemini_model_complete(

lightrag/llm/openai.py

@@ -174,6 +174,7 @@ async def openai_complete_if_cache(
                 explicit parameters (api_key, base_url).
             - hashing_kv: Will be removed from kwargs before passing to OpenAI.
             - keyword_extraction: Will be removed from kwargs before passing to OpenAI.
+            - stream: Whether to stream the response. Default is False.
 
     Returns:
         The completed text (with integrated COT content if available) or an async iterator