Fix tool-call fallback for structured completion content

.semversioner/next-release/patch-20260312083316079043.json

@@ -0,0 +1,4 @@
+{
+  "type": "patch",
+  "description": "Fix structured response parsing when providers return tool calls with empty message content"
+}

packages/graphrag-llm/graphrag_llm/types/types.py

@@ -97,8 +97,27 @@ class LLMCompletionResponse(ChatCompletion, Generic[ResponseFormat]):
     @computed_field
     @property
     def content(self) -> str:
-        """Get the content of the first choice message."""
-        return self.choices[0].message.content or ""
+        """Get the content of the first choice message.
+
+        Falls back to function tool-call arguments when `content` is empty.
+        Some providers return structured payloads via tool calls instead of
+        assistant message content.
+        """
+        if not self.choices:
+            return ""
+
+        message = self.choices[0].message
+        if message.content:
+            return message.content
+
+        if message.tool_calls:
+            for tool_call in message.tool_calls:
+                if tool_call.type != "function":
+                    continue
+                if tool_call.function.arguments:
+                    return tool_call.function.arguments
+
+        return ""
 
 
 class LLMCompletionArgs(

packages/graphrag-llm/graphrag_llm/utils/gather_completion_response.py

@@ -30,7 +30,7 @@ def gather_completion_response(
     if isinstance(response, Iterator):
         return "".join(chunk.choices[0].delta.content or "" for chunk in response)
 
-    return response.choices[0].message.content or ""
+    return response.content
 
 
 async def gather_completion_response_async(
@@ -54,4 +54,4 @@ async def gather_completion_response_async(
 
         return gathered_content
 
-    return response.choices[0].message.content or ""
+    return response.content

tests/unit/utils/test_completion_response.py

@@ -0,0 +1,106 @@
+# Copyright (c) 2024 Microsoft Corporation.
+# Licensed under the MIT License
+
+import asyncio
+
+from graphrag_llm.types import LLMCompletionResponse
+from graphrag_llm.utils import (
+    gather_completion_response,
+    gather_completion_response_async,
+    structure_completion_response,
+)
+from pydantic import BaseModel
+
+
+class RatingResponse(BaseModel):
+    rating: int
+
+
+def _create_completion_response(
+    *,
+    content: str | None,
+    tool_call_arguments: str | None = None,
+) -> LLMCompletionResponse:
+    message: dict = {
+        "role": "assistant",
+        "content": content,
+    }
+
+    if tool_call_arguments is not None:
+        message["tool_calls"] = [
+            {
+                "id": "call_1",
+                "type": "function",
+                "function": {
+                    "name": "structured_output",
+                    "arguments": tool_call_arguments,
+                },
+            }
+        ]
+
+    return LLMCompletionResponse(
+        id="completion-id",
+        object="chat.completion",
+        created=0,
+        model="mock-model",
+        choices=[
+            {
+                "index": 0,
+                "message": message,
+                "finish_reason": "stop",
+            }
+        ],
+        usage={
+            "prompt_tokens": 0,
+            "completion_tokens": 0,
+            "total_tokens": 0,
+        },
+        formatted_response=None,
+    )
+
+
+def test_content_prefers_message_content() -> None:
+    response = _create_completion_response(
+        content="plain text",
+        tool_call_arguments='{"rating": 9}',
+    )
+
+    assert response.content == "plain text"
+
+
+def test_content_falls_back_to_function_tool_call_arguments() -> None:
+    response = _create_completion_response(
+        content=None,
+        tool_call_arguments='{"rating": 7}',
+    )
+
+    assert response.content == '{"rating": 7}'
+
+
+def test_gather_completion_response_falls_back_to_tool_call_arguments() -> None:
+    response = _create_completion_response(
+        content=None,
+        tool_call_arguments='{"rating": 3}',
+    )
+
+    assert gather_completion_response(response) == '{"rating": 3}'
+
+
+def test_gather_completion_response_async_falls_back_to_tool_call_arguments() -> None:
+    response = _create_completion_response(
+        content=None,
+        tool_call_arguments='{"rating": 5}',
+    )
+
+    gathered_response = asyncio.run(gather_completion_response_async(response))
+    assert gathered_response == '{"rating": 5}'
+
+
+def test_structure_completion_response_uses_tool_call_arguments() -> None:
+    response = _create_completion_response(
+        content=None,
+        tool_call_arguments='{"rating": 11}',
+    )
+
+    parsed = structure_completion_response(response.content, RatingResponse)
+    assert parsed.rating == 11