From 2aabd1d90ebce52fff4c855711671ad42a9adbfd Mon Sep 17 00:00:00 2001
From: Will Miao <willmiao.dev@gmail.com>
Date: Thu, 23 Jul 2026 09:17:29 +0800
Subject: [PATCH] fix(ai): use json_schema instead of json_object for broader
 provider compatibility (#1033)

LM Studio and some other OpenAI-compatible servers reject
response_format=json_object but accept json_schema. Switch to the
equivalent json_schema format and add a fallback that retries
without response_format when the provider rejects the format type.
---
 py/services/llm_service.py         | 50 +++++++++++++++++++++++++-----
 tests/services/test_llm_service.py | 50 ++++++++++++++++++++++++++++++
 2 files changed, 92 insertions(+), 8 deletions(-)

diff --git a/py/services/llm_service.py b/py/services/llm_service.py
index b6854302..ea4e83dc 100644
--- a/py/services/llm_service.py
+++ b/py/services/llm_service.py
@@ -566,18 +566,52 @@ class LLMService:
         if effective_max is None:
             effective_max = 4096
 
-        result = await self.chat_completion(
-            messages=messages,
-            model=model,
-            temperature=temperature,
-            response_format={"type": "json_object"},
-            max_tokens=effective_max,
-        )
+        # Use json_schema (not json_object) for broader provider compatibility:
+        # LM Studio and some other OpenAI-compatible servers reject
+        # json_object but accept json_schema.  {"type": "object"} is
+        # functionally equivalent — it accepts any JSON object without
+        # constraining specific fields.
+        response_format = {
+            "type": "json_schema",
+            "json_schema": {
+                "name": "metadata",
+                "schema": {"type": "object"},
+            },
+        }
+
+        try:
+            result = await self.chat_completion(
+                messages=messages,
+                model=model,
+                temperature=temperature,
+                response_format=response_format,
+                max_tokens=effective_max,
+            )
+        except LLMResponseError as e:
+            # Only fall back when the provider rejects the response_format
+            # type value (e.g. "'response_format.type' must be...").  Avoid
+            # catching unrelated 400 errors whose body happens to mention
+            # "response_format" (e.g. "model does not support
+            # response_format restrictions on this endpoint").
+            if "'response_format.type'" not in str(e).lower():
+                raise
+            logger.info(
+                "Provider rejected response_format, retrying without it. "
+                "Falling back to prompt-only JSON mode. Error: %s",
+                e,
+            )
+            result = await self.chat_completion(
+                messages=messages,
+                model=model,
+                temperature=temperature,
+                response_format=None,
+                max_tokens=effective_max,
+            )
 
         content = result.get("content", "") or ""
         if not content:
             raise LLMResponseError(
-                "LLM returned empty content in json_object mode. "
+                "LLM returned empty content. "
                 f"Raw response: {json.dumps(result)[:500]}"
             )
 
diff --git a/tests/services/test_llm_service.py b/tests/services/test_llm_service.py
index 5fb1a5e1..c4d38fe8 100644
--- a/tests/services/test_llm_service.py
+++ b/tests/services/test_llm_service.py
@@ -243,6 +243,56 @@ class TestLLMServiceChatCompletionJson:
 
         assert result == {"key": "value"}
 
+    @pytest.mark.asyncio
+    async def test_chat_completion_json_falls_back_on_response_format_rejection(
+        self, llm_service,
+    ):
+        """Retry without response_format when provider rejects it (HTTP 400)."""
+        error_response = MockResponse(
+            400,
+            text_data=(
+                '{"error":"\'response_format.type\' must be '
+                '\'json_schema\' or \'text\'"}'
+            ),
+        )
+        success_response = MockResponse(
+            200,
+            json_data={
+                "choices": [{"message": {"content": '{"key": "value"}'}}],
+                "usage": {},
+                "model": "local-model",
+            },
+        )
+
+        call_index = 0
+
+        class FallbackMockSession:
+            def __init__(self):
+                self.last_url = None
+                self.last_json = None
+
+            def post(self, url, json=None, headers=None):
+                nonlocal call_index
+                self.last_url = url
+                self.last_json = json
+                call_index += 1
+                return error_response if call_index == 1 else success_response
+
+            async def __aenter__(self):
+                return self
+
+            async def __aexit__(self, *args):
+                pass
+
+        with mock.patch("aiohttp.ClientSession", return_value=FallbackMockSession()):
+            result = await llm_service.chat_completion_json(
+                system_prompt="You are helpful.",
+                user_prompt="Return JSON.",
+            )
+
+        assert result == {"key": "value"}
+        assert call_index == 2
+
     @pytest.mark.asyncio
     async def test_chat_completion_json_raises_on_non_json(self, llm_service):
         # Non-JSON content raises LLMResponseError (salvage also fails)