fix(openai): preserve refusal text in traces

langfuse/openai.py

@@ -475,6 +475,9 @@ def _extract_chat_response(kwargs: Any) -> Any:
     if kwargs.get("tool_calls") is not None:
         response.update({"tool_calls": kwargs["tool_calls"]})
 
+    if kwargs.get("refusal") is not None:
+        response.update({"refusal": kwargs["refusal"]})
+
     if kwargs.get("audio") is not None:
         audio = kwargs["audio"].__dict__
 
@@ -827,12 +830,9 @@ def _extract_streamed_openai_response(resource: Any, chunks: Any) -> Any:
                 if delta.get("role", None) is not None:
                     completion["role"] = delta["role"]
 
-                if delta.get("content", None) is not None:
-                    completion["content"] = (
-                        delta.get("content", None)
-                        if completion["content"] is None
-                        else completion["content"] + delta.get("content", None)
-                    )
+                for field in ("content", "refusal"):
+                    if delta.get(field) is not None:
+                        completion[field] = (completion[field] or "") + delta[field]
 
                 if delta.get("function_call", None) is not None:
                     curr = completion["function_call"]
@@ -907,6 +907,9 @@ def _extract_streamed_openai_response(resource: Any, chunks: Any) -> Any:
     def get_response_for_chat() -> Any:
         content = completion["content"]
 
+        if completion["refusal"] is not None:
+            return _extract_chat_response({**completion, "role": "assistant"})
+
         if completion["tool_calls"]:
             response = {
                 "role": "assistant",

tests/unit/test_openai.py

@@ -1,7 +1,9 @@
 import asyncio
+import json
 from types import SimpleNamespace
 from unittest.mock import patch
 
+import httpx
 import pytest
 from openai.types.responses import ParsedResponseOutputMessage, ParsedResponseOutputText
 from pydantic import BaseModel
@@ -1367,6 +1369,95 @@ def handler(request: httpx.Request) -> httpx.Response:
     )
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize("async_client", [False, True])
+@pytest.mark.parametrize("stream", [False, True])
+async def test_chat_completion_captures_refusal(
+    langfuse_memory_client, get_span, json_attr, async_client, stream
+):
+    refusal = "I cannot help with that request."
+    payload = _chat_completion_payload()
+    payload["choices"][0]["message"] = {
+        "role": "assistant",
+        "content": None,
+        "refusal": refusal,
+    }
+
+    def handler(request):
+        if not stream:
+            return httpx.Response(200, json=payload)
+
+        events = []
+        for part in [None, "I cannot help ", "with that request."]:
+            events.append(
+                {
+                    "id": payload["id"],
+                    "object": "chat.completion.chunk",
+                    "created": payload["created"],
+                    "model": payload["model"],
+                    "choices": [
+                        {
+                            "index": 0,
+                            "delta": {"role": "assistant", "refusal": part},
+                            "finish_reason": None,
+                        }
+                    ],
+                }
+            )
+        events.append(
+            {
+                **events[-1],
+                "choices": [{"index": 0, "delta": {}, "finish_reason": "stop"}],
+                "usage": payload["usage"],
+            }
+        )
+        body = "".join(f"data: {json.dumps(event)}\n\n" for event in events)
+        return httpx.Response(
+            200,
+            content=body + "data: [DONE]\n\n",
+            headers={"content-type": "text/event-stream"},
+        )
+
+    kwargs = {
+        "model": "gpt-4o-mini",
+        "messages": [{"role": "user", "content": "test request"}],
+        "stream": stream,
+    }
+    if async_client:
+        async with lf_openai.AsyncOpenAI(
+            api_key="test",
+            http_client=httpx.AsyncClient(transport=httpx.MockTransport(handler)),
+        ) as client:
+            result = await client.chat.completions.create(**kwargs)
+            if stream:
+                result = [chunk async for chunk in result]
+    else:
+        with lf_openai.OpenAI(
+            api_key="test",
+            http_client=httpx.Client(transport=httpx.MockTransport(handler)),
+        ) as client:
+            result = client.chat.completions.create(**kwargs)
+            if stream:
+                result = list(result)
+
+    if stream:
+        assert (
+            "".join(chunk.choices[0].delta.refusal or "" for chunk in result) == refusal
+        )
+    else:
+        assert result.choices[0].message.refusal == refusal
+
+    langfuse_memory_client.flush()
+    span = get_span("OpenAI-generation")
+    assert json_attr(span, LangfuseOtelSpanAttributes.OBSERVATION_OUTPUT) == {
+        "role": "assistant",
+        "content": None,
+        "refusal": refusal,
+    }
+    usage = json_attr(span, LangfuseOtelSpanAttributes.OBSERVATION_USAGE_DETAILS)
+    assert usage["total_tokens"] == payload["usage"]["total_tokens"]
+
+
 def test_with_raw_response_chat_completion_captures_output_and_usage(
     langfuse_memory_client, get_span, json_attr
 ):