🐛 The prompt includes the thought process #2038

Author: Phinease

backend/utils/llm_utils.py

@@ -18,16 +18,37 @@ def _process_thinking_tokens(
 ) -> bool:
     """
     Process tokens to filter out thinking content between <think> and </think> tags.
+    Handles cases where providers only send a closing tag or mix reasoning_content.
     """
-    if is_thinking:
-        return THINK_END_PATTERN not in new_token
+    # Check for end tag first, as it might appear in the same token as start tag
+    if THINK_END_PATTERN in new_token:
+        # If we were never in think mode, treat everything accumulated so far as reasoning and clear it
+        if not is_thinking:
+            token_join.clear()
+            if callback:
+                callback("")  # clear any previously streamed reasoning content
+
+        # Exit thinking mode and only keep content after </think>
+        _, _, after_end = new_token.partition(THINK_END_PATTERN)
+        is_thinking = False
+        new_token = after_end
+        # Continue processing the remaining content in this token
 
+    # Check for start tag (after processing end tag, in case both are in the same token)
     if THINK_START_PATTERN in new_token:
+        # Drop any content before <think> and switch to thinking mode
+        _, _, after_start = new_token.partition(THINK_START_PATTERN)
+        new_token = after_start
+        is_thinking = True
+
+    if is_thinking:
+        # Still inside thinking content; ignore until we exit
         return True
 
-    token_join.append(new_token)
-    if callback:
-        callback("".join(token_join))
+    if new_token:
+        token_join.append(new_token)
+        if callback:
+            callback("".join(token_join))
 
     return False
 
@@ -46,8 +67,8 @@ def call_llm_for_system_prompt(
 
     llm = OpenAIModel(
         model_id=get_model_name_from_config(llm_model_config) if llm_model_config else "",
-        api_base=llm_model_config.get("base_url", ""),
-        api_key=llm_model_config.get("api_key", ""),
+        api_base=llm_model_config.get("base_url", "") if llm_model_config else "",
+        api_key=llm_model_config.get("api_key", "") if llm_model_config else "",
         temperature=0.3,
         top_p=0.95,
     )
@@ -65,16 +86,38 @@ def call_llm_for_system_prompt(
         current_request = llm.client.chat.completions.create(stream=True, **completion_kwargs)
         token_join: List[str] = []
         is_thinking = False
+        reasoning_content_seen = False
+        content_tokens_seen = 0
         for chunk in current_request:
-            new_token = chunk.choices[0].delta.content
+            delta = chunk.choices[0].delta
+            reasoning_content = getattr(delta, "reasoning_content", None)
+            new_token = delta.content
+
+            # Note: reasoning_content is separate metadata and doesn't affect content filtering
+            # We only filter content based on <think> tags in delta.content
+            if reasoning_content:
+                reasoning_content_seen = True
+                logger.debug("Received reasoning_content (metadata only, not filtering content)")
+
+            # Process content token if it exists
             if new_token is not None:
+                content_tokens_seen += 1
                 is_thinking = _process_thinking_tokens(
                     new_token,
                     is_thinking,
                     token_join,
                     callback,
                 )
-        return "".join(token_join)
+        
+        result = "".join(token_join)
+        if not result and content_tokens_seen > 0:
+            logger.warning(
+                "Generated prompt is empty but %d content tokens were processed. "
+                "This suggests all content was filtered out.",
+                content_tokens_seen
+            )
+        
+        return result
     except Exception as exc:
         logger.error("Failed to generate prompt from LLM: %s", str(exc))
         raise