Merge 28ecd2e7f9 into 2c919efa69

api/core/app/apps/base_app_runner.py

@@ -247,6 +247,7 @@ class AppRunner:
         model: str = ""
         prompt_messages: list[PromptMessage] = []
         text = ""
+        reasoning_content = ""
         usage = None
         for result in invoke_result:
             if not agent:
@@ -266,6 +267,10 @@ class AppRunner:
                     else:
                         text += content  # failback to str
 
+            # Handle reasoning content from delta (e.g., Ollama's thinking field)
+            if result.delta.reasoning_content:
+                reasoning_content += result.delta.reasoning_content
+
             if not model:
                 model = result.model
 
@@ -279,7 +284,11 @@ class AppRunner:
             usage = LLMUsage.empty_usage()
 
         llm_result = LLMResult(
-            model=model, prompt_messages=prompt_messages, message=AssistantPromptMessage(content=text), usage=usage
+            model=model,
+            prompt_messages=prompt_messages,
+            message=AssistantPromptMessage(content=text),
+            usage=usage,
+            reasoning_content=reasoning_content or None,
         )
 
         queue_manager.publish(

api/core/model_runtime/entities/llm_entities.py

@@ -192,6 +192,7 @@ class LLMResultChunkDelta(BaseModel):
     message: AssistantPromptMessage
     usage: LLMUsage | None = None
     finish_reason: str | None = None
+    reasoning_content: str | None = None
 
 
 class LLMResultChunk(BaseModel):

api/core/plugin/impl/base.py

@@ -9,6 +9,7 @@ from pydantic import BaseModel
 from yarl import URL
 
 from configs import dify_config
+from core.model_runtime.entities.llm_entities import LLMResultChunk
 from core.model_runtime.errors.invoke import (
     InvokeAuthorizationError,
     InvokeBadRequestError,
@@ -255,7 +256,17 @@ class BasePluginClient:
         """
         for line in self._stream_request(method, path, params, headers, data, files):
             try:
-                rep = PluginDaemonBasicResponse[type_].model_validate_json(line)  # type: ignore
+                line_data = json.loads(line)
+
+                if isinstance(line_data, dict) and type_ is LLMResultChunk:
+                    if "data" in line_data and isinstance(line_data["data"], dict):
+                        data_dict = line_data["data"]
+                        if "delta" in data_dict and isinstance(data_dict["delta"], dict):
+                            delta_dict = data_dict["delta"]
+                            if "thinking" in delta_dict and "reasoning_content" not in delta_dict:
+                                delta_dict["reasoning_content"] = delta_dict.pop("thinking")
+
+                rep = PluginDaemonBasicResponse[type_].model_validate(line_data)  # type: ignore
             except (ValueError, TypeError):
                 # TODO modify this when line_data has code and message
                 try:

api/core/workflow/nodes/llm/node.py

@@ -444,6 +444,7 @@ class LLMNode(Node[LLMNodeData]):
         usage = LLMUsage.empty_usage()
         finish_reason = None
         full_text_buffer = io.StringIO()
+        reasoning_content_buffer = io.StringIO()
 
         # Initialize streaming metrics tracking
         start_time = request_start_time if request_start_time is not None else time.perf_counter()
@@ -478,6 +479,15 @@ class LLMNode(Node[LLMNodeData]):
                             is_final=False,
                         )
 
+                    # Handle reasoning content from delta (e.g., Ollama's thinking field)
+                    if result.delta.reasoning_content:
+                        reasoning_content_buffer.write(result.delta.reasoning_content)
+                        yield StreamChunkEvent(
+                            selector=[node_id, "reasoning_content"],
+                            chunk=result.delta.reasoning_content,
+                            is_final=False,
+                        )
+
                     # Update the whole metadata
                     if not model and result.model:
                         model = result.model
@@ -494,6 +504,7 @@ class LLMNode(Node[LLMNodeData]):
 
         # Extract reasoning content from <think> tags in the main text
         full_text = full_text_buffer.getvalue()
+        streamed_reasoning_content = reasoning_content_buffer.getvalue()
 
         if reasoning_format == "tagged":
             # Keep <think> tags in text for backward compatibility
@@ -501,7 +512,8 @@ class LLMNode(Node[LLMNodeData]):
             reasoning_content = ""
         else:
             # Extract clean text and reasoning from <think> tags
-            clean_text, reasoning_content = LLMNode._split_reasoning(full_text, reasoning_format)
+            clean_text, extracted_reasoning_content = LLMNode._split_reasoning(full_text, reasoning_format)
+            reasoning_content = streamed_reasoning_content or extracted_reasoning_content
 
         # Calculate streaming metrics
         end_time = time.perf_counter()
@@ -1158,7 +1170,8 @@ class LLMNode(Node[LLMNodeData]):
             reasoning_content = ""
         else:
             # Extract clean text and reasoning from <think> tags
-            clean_text, reasoning_content = LLMNode._split_reasoning(full_text, reasoning_format)
+            clean_text, extracted_reasoning_content = LLMNode._split_reasoning(full_text, reasoning_format)
+            reasoning_content = invoke_result.reasoning_content or extracted_reasoning_content
 
         event = ModelInvokeCompletedEvent(
             # Use clean_text for separated mode, full_text for tagged mode