fix(openai): tool_calls dropped when content chunk precedes tool deltas in stream

langfuse/openai.py

@@ -729,27 +729,30 @@ def _extract_streamed_openai_response(resource: Any, chunks: Any) -> Any:
                 completion += choice.get("text", "")
 
     def get_response_for_chat() -> Any:
-        return (
-            completion["content"]
-            or (
-                completion["function_call"]
-                and {
-                    "role": "assistant",
-                    "function_call": completion["function_call"],
-                }
-            )
-            or (
-                completion["tool_calls"]
-                and {
-                    "role": "assistant",
-                    # "tool_calls": [{"function": completion["tool_calls"]}],
-                    "tool_calls": [
-                        {"function": data} for data in completion["tool_calls"]
-                    ],
-                }
-            )
-            or None
-        )
+        # tool_calls must be checked before content: models like Qwen/DeepSeek emit
+        # whitespace content chunks (e.g. "\n\n") before streaming tool call deltas,
+        # which would otherwise cause the content branch to short-circuit and silently
+        # drop the collected tool_calls.
+        if completion["tool_calls"]:
+            result: Any = {
+                "role": "assistant",
+                "tool_calls": [
+                    {"function": data} for data in completion["tool_calls"]
+                ],
+            }
+            # Preserve non-whitespace content that co-exists with tool calls
+            # (some reasoning models emit a brief text preamble before calling tools).
+            if completion["content"] and completion["content"].strip():
+                result["content"] = completion["content"]
+            return result
+
+        if completion["function_call"]:
+            return {
+                "role": "assistant",
+                "function_call": completion["function_call"],
+            }
+
+        return completion["content"] or None
 
     return (
         model,

tests/test_openai_streaming_unit.py

@@ -0,0 +1,255 @@
+"""Unit tests for _extract_streamed_openai_response / get_response_for_chat.
+
+Covers the bug where a non-empty content chunk (e.g. "\n\n") emitted before
+tool-call deltas caused get_response_for_chat() to short-circuit and silently
+drop all collected tool_calls from the logged generation output.
+
+No real OpenAI API calls — chunks are built from SimpleNamespace objects that
+mirror the __dict__ structure of openai-python v1 Pydantic models.
+"""
+
+import types
+from dataclasses import dataclass
+from typing import Any, List, Optional
+from unittest.mock import patch
+
+import pytest
+
+from langfuse.openai import OpenAiDefinition, _extract_streamed_openai_response
+
+
+# ---------------------------------------------------------------------------
+# Helpers: fake OpenAI v1 streaming chunk objects
+# ---------------------------------------------------------------------------
+
+def _make_tool_call_delta(
+    name: Optional[str] = None,
+    arguments: str = "",
+    index: int = 0,
+    tool_id: Optional[str] = None,
+    call_type: Optional[str] = None,
+) -> Any:
+    """Build a ChoiceDeltaToolCall-alike SimpleNamespace."""
+    function = types.SimpleNamespace(name=name, arguments=arguments)
+    return types.SimpleNamespace(
+        index=index,
+        id=tool_id,
+        type=call_type,
+        function=function,
+    )
+
+
+def _make_chunk(
+    content: Optional[str] = None,
+    tool_calls: Optional[List[Any]] = None,
+    function_call: Any = None,
+    role: Optional[str] = None,
+    finish_reason: Optional[str] = None,
+    model: str = "gpt-4o",
+) -> Any:
+    """Build a streaming chunk SimpleNamespace (mirrors chunk.__dict__ in v1)."""
+    delta = types.SimpleNamespace(
+        role=role,
+        content=content,
+        tool_calls=tool_calls,
+        function_call=function_call,
+    )
+    choice = types.SimpleNamespace(delta=delta, finish_reason=finish_reason)
+    return types.SimpleNamespace(model=model, choices=[choice], usage=None)
+
+
+def _chat_resource() -> OpenAiDefinition:
+    return OpenAiDefinition(
+        module="openai",
+        object="ChatCompletion",
+        method="create",
+        type="chat",
+        sync=True,
+    )
+
+
+def _run(chunks: List[Any]) -> Any:
+    """Run _extract_streamed_openai_response with is_openai_v1 patched to True."""
+    with patch("langfuse.openai._is_openai_v1", return_value=True):
+        _, response, _, _ = _extract_streamed_openai_response(_chat_resource(), iter(chunks))
+    return response
+
+
+# ---------------------------------------------------------------------------
+# Bug reproduction: content chunk before tool_calls
+# ---------------------------------------------------------------------------
+
+
+class TestToolCallsWithPrecedingContentChunk:
+    """
+    Models like Qwen/DeepSeek sometimes emit a whitespace content chunk
+    (e.g. "\n\n") before beginning to stream tool-call deltas. Previously
+    get_response_for_chat() evaluated `completion["content"] or ...` and
+    returned the content string immediately, dropping the tool_calls entirely.
+    """
+
+    def test_tool_calls_not_dropped_when_whitespace_content_precedes_them(self):
+        chunks = [
+            _make_chunk(role="assistant"),
+            _make_chunk(content="\n\n"),   # spurious whitespace before tool call
+            _make_chunk(
+                tool_calls=[_make_tool_call_delta(name="get_weather", arguments="")],
+            ),
+            _make_chunk(
+                tool_calls=[_make_tool_call_delta(name=None, arguments='{"city": "Paris"}')],
+            ),
+            _make_chunk(finish_reason="tool_calls"),
+        ]
+        result = _run(chunks)
+
+        assert isinstance(result, dict), "Expected a dict, not a plain string"
+        assert "tool_calls" in result, "tool_calls must not be dropped"
+        assert result["tool_calls"][0]["function"]["name"] == "get_weather"
+        assert result["tool_calls"][0]["function"]["arguments"] == '{"city": "Paris"}'
+
+    def test_whitespace_only_content_not_included_in_result(self):
+        """A leading "\n\n" is whitespace-only and should be omitted from output."""
+        chunks = [
+            _make_chunk(role="assistant"),
+            _make_chunk(content="\n\n"),
+            _make_chunk(
+                tool_calls=[_make_tool_call_delta(name="search", arguments='{"q":"hi"}')],
+            ),
+            _make_chunk(finish_reason="tool_calls"),
+        ]
+        result = _run(chunks)
+
+        assert "content" not in result or result.get("content") is None
+
+    def test_meaningful_content_preserved_alongside_tool_calls(self):
+        """When content has real text (not just whitespace), it should be kept."""
+        chunks = [
+            _make_chunk(role="assistant"),
+            _make_chunk(content="Sure, let me check that. "),
+            _make_chunk(
+                tool_calls=[_make_tool_call_delta(name="lookup", arguments='{"id":1}')],
+            ),
+            _make_chunk(finish_reason="tool_calls"),
+        ]
+        result = _run(chunks)
+
+        assert "tool_calls" in result
+        assert result.get("content") == "Sure, let me check that. "
+
+    def test_non_whitespace_content_before_tool_calls_preserves_both(self):
+        chunks = [
+            _make_chunk(role="assistant"),
+            _make_chunk(content="I'll call"),
+            _make_chunk(content=" the tool."),
+            _make_chunk(
+                tool_calls=[_make_tool_call_delta(name="do_thing", arguments="{}")],
+            ),
+            _make_chunk(finish_reason="tool_calls"),
+        ]
+        result = _run(chunks)
+
+        assert result["tool_calls"][0]["function"]["name"] == "do_thing"
+        assert result.get("content") == "I'll call the tool."
+
+
+# ---------------------------------------------------------------------------
+# Baseline: pure content response (no tools)
+# ---------------------------------------------------------------------------
+
+
+class TestPureContentResponse:
+    def test_plain_text_response_returned_as_string(self):
+        chunks = [
+            _make_chunk(role="assistant"),
+            _make_chunk(content="Hello, "),
+            _make_chunk(content="world!"),
+            _make_chunk(finish_reason="stop"),
+        ]
+        result = _run(chunks)
+        assert result == "Hello, world!"
+
+    def test_empty_stream_returns_none(self):
+        result = _run([])
+        assert result is None
+
+
+# ---------------------------------------------------------------------------
+# Pure tool-call response (no content at all)
+# ---------------------------------------------------------------------------
+
+
+class TestPureToolCallResponse:
+    def test_tool_calls_returned_without_content(self):
+        chunks = [
+            _make_chunk(role="assistant"),
+            _make_chunk(
+                tool_calls=[_make_tool_call_delta(name="get_price", arguments='{"sku":"A1"}')],
+            ),
+            _make_chunk(finish_reason="tool_calls"),
+        ]
+        result = _run(chunks)
+
+        assert isinstance(result, dict)
+        assert "tool_calls" in result
+        assert result["tool_calls"][0]["function"]["name"] == "get_price"
+        assert "content" not in result
+
+    def test_multiple_tool_calls_all_returned(self):
+        chunks = [
+            _make_chunk(role="assistant"),
+            _make_chunk(
+                tool_calls=[_make_tool_call_delta(name="tool_a", arguments='{"x":1}')],
+            ),
+            # second tool call — name triggers a new entry in the accumulator
+            _make_chunk(
+                tool_calls=[_make_tool_call_delta(name="tool_b", arguments='{"y":2}')],
+            ),
+            _make_chunk(finish_reason="tool_calls"),
+        ]
+        result = _run(chunks)
+
+        assert len(result["tool_calls"]) == 2
+        names = {tc["function"]["name"] for tc in result["tool_calls"]}
+        assert names == {"tool_a", "tool_b"}
+
+
+# ---------------------------------------------------------------------------
+# Legacy function_call (OpenAI v0 format)
+# ---------------------------------------------------------------------------
+
+
+class TestFunctionCallResponse:
+    def test_function_call_returned_when_no_tool_calls(self):
+        # function_call uses a different delta key path; simulate with direct
+        # injection via a SimpleNamespace that has function_call set
+        chunks = [
+            _make_chunk(role="assistant"),
+        ]
+        # Patch the completion dict after the fact is tricky; instead, build
+        # a chunk that triggers the function_call accumulation path.
+        fn_chunk = types.SimpleNamespace(
+            model="gpt-3.5-turbo",
+            choices=[
+                types.SimpleNamespace(
+                    delta=types.SimpleNamespace(
+                        role=None,
+                        content=None,
+                        tool_calls=None,
+                        function_call=types.SimpleNamespace(
+                            name="old_fn",
+                            arguments='{"a":1}',
+                        ),
+                    ),
+                    finish_reason=None,
+                )
+            ],
+            usage=None,
+        )
+        with patch("langfuse.openai._is_openai_v1", return_value=True):
+            _, result, _, _ = _extract_streamed_openai_response(
+                _chat_resource(), iter([fn_chunk])
+            )
+
+        assert isinstance(result, dict)
+        assert "function_call" in result
+        assert result["function_call"]["name"] == "old_fn"