From f55faae31b0e9e4b308056c6ad956737f3abd7a0 Mon Sep 17 00:00:00 2001
From: Novice <novice12185727@gmail.com>
Date: Thu, 25 Dec 2025 13:59:38 +0800
Subject: [PATCH] chore: strip reasoning from chatflow answers and persist
 generation details

---
 .../advanced_chat/generate_task_pipeline.py   | 18 ++++++++++++-
 .../easy_ui_based_generate_task_pipeline.py   | 25 ++++++++++++++++---
 2 files changed, 38 insertions(+), 5 deletions(-)

diff --git a/api/core/app/apps/advanced_chat/generate_task_pipeline.py b/api/core/app/apps/advanced_chat/generate_task_pipeline.py
index 0d5e1e5dfd..53fa27cca7 100644
--- a/api/core/app/apps/advanced_chat/generate_task_pipeline.py
+++ b/api/core/app/apps/advanced_chat/generate_task_pipeline.py
@@ -529,7 +529,9 @@ class AdvancedChatAppGenerateTaskPipeline(GraphRuntimeStateSupport):
         match chunk_type:
             case ChunkType.TEXT:
                 self._stream_buffer.record_text_chunk(delta_text)
+                self._task_state.answer += delta_text
             case ChunkType.THOUGHT:
+                # Reasoning should not be part of final answer text
                 self._stream_buffer.record_thought_chunk(delta_text)
             case ChunkType.TOOL_CALL:
                 self._stream_buffer.record_tool_call(
@@ -542,8 +544,8 @@ class AdvancedChatAppGenerateTaskPipeline(GraphRuntimeStateSupport):
                     tool_call_id=tool_call_id,
                     result=delta_text,
                 )
+                self._task_state.answer += delta_text
 
-        self._task_state.answer += delta_text
         yield self._message_cycle_manager.message_to_stream_response(
             answer=delta_text,
             message_id=self._message_id,
@@ -920,6 +922,7 @@ class AdvancedChatAppGenerateTaskPipeline(GraphRuntimeStateSupport):
 
         # If there are assistant files, remove markdown image links from answer
         answer_text = self._task_state.answer
+        answer_text = self._strip_think_blocks(answer_text)
         if self._recorded_files:
             # Remove markdown image links since we're storing files separately
             answer_text = re.sub(r"!\[.*?\]\(.*?\)", "", answer_text).strip()
@@ -971,6 +974,19 @@ class AdvancedChatAppGenerateTaskPipeline(GraphRuntimeStateSupport):
         ]
         session.add_all(message_files)
 
+        # Save generation detail (reasoning/tool calls/sequence) from stream buffer
+        self._save_generation_detail(session=session, message=message)
+
+    @staticmethod
+    def _strip_think_blocks(text: str) -> str:
+        """Remove <think>...</think> blocks (including their content) from text."""
+        if not text or "<think" not in text.lower():
+            return text
+
+        clean_text = re.sub(r"<think[^>]*>.*?</think>", "", text, flags=re.IGNORECASE | re.DOTALL)
+        clean_text = re.sub(r"\n\s*\n", "\n\n", clean_text).strip()
+        return clean_text
+
     def _save_generation_detail(self, *, session: Session, message: Message) -> None:
         """
         Save LLM generation detail for Chatflow using stream event buffer.
diff --git a/api/core/app/task_pipeline/easy_ui_based_generate_task_pipeline.py b/api/core/app/task_pipeline/easy_ui_based_generate_task_pipeline.py
index 6cbd48e27b..c4ea428270 100644
--- a/api/core/app/task_pipeline/easy_ui_based_generate_task_pipeline.py
+++ b/api/core/app/task_pipeline/easy_ui_based_generate_task_pipeline.py
@@ -441,6 +441,7 @@ class EasyUIBasedGenerateTaskPipeline(BasedGenerateTaskPipeline):
         if agent_thoughts:
             # Agent-Chat mode: merge MessageAgentThought records
             content_pos = 0
+            cleaned_answer_parts: list[str] = []
             for thought in agent_thoughts:
                 # Add thought/reasoning
                 if thought.thought:
@@ -466,10 +467,26 @@ class EasyUIBasedGenerateTaskPipeline(BasedGenerateTaskPipeline):
 
                 # Add answer content if present
                 if thought.answer:
-                    start = content_pos
-                    end = content_pos + len(thought.answer)
-                    sequence.append({"type": "content", "start": start, "end": end})
-                    content_pos = end
+                    content_text = thought.answer
+                    if "<think" in content_text.lower():
+                        clean_answer, extracted_reasoning = self._split_reasoning_from_answer(content_text)
+                        if extracted_reasoning:
+                            reasoning_list.append(extracted_reasoning)
+                            sequence.append({"type": "reasoning", "index": len(reasoning_list) - 1})
+                        content_text = clean_answer
+                        thought.answer = clean_answer or content_text
+
+                    if content_text:
+                        start = content_pos
+                        end = content_pos + len(content_text)
+                        sequence.append({"type": "content", "start": start, "end": end})
+                        content_pos = end
+                        cleaned_answer_parts.append(content_text)
+
+            if cleaned_answer_parts:
+                merged_answer = "".join(cleaned_answer_parts)
+                message.answer = merged_answer
+                llm_result.message.content = merged_answer
         else:
             # Completion/Chat mode: use reasoning_content from llm_result
             reasoning_content = llm_result.reasoning_content