fix(llm): drop falsy logprobs from payloads for strict providers

openrag/services/inference/vllm_client.py

@@ -115,6 +115,22 @@ def _log_safe_error_detail(exc: BaseException) -> dict:
 # ---------------------------------------------------------------------------
 
 
+def _strip_falsy_logprobs(payload: dict) -> dict:
+    """Drop a falsy ``logprobs`` (and its dependent ``top_logprobs``) from *payload*.
+
+    ``logprobs: false`` is already the OpenAI default, so sending it adds
+    nothing — but strict providers whose schema lacks the field reject it by
+    name whatever the value, e.g. Gemini. Without
+    this, the config default (``LLMParamsConfig.logprobs = False``) lands in
+    ``self._defaults`` and is sent on every request. A truthy value is a
+    deliberate opt-in and is forwarded as-is.
+    """
+    if not payload.get("logprobs"):
+        payload.pop("logprobs", None)
+        payload.pop("top_logprobs", None)
+    return payload
+
+
 @llm_registry.register("vllm")
 class VLLMClient(LLM):
     """OpenAI-compatible LLM client backed by vLLM.
@@ -192,6 +208,7 @@ def _chat_payload_kwargs(self, kwargs: dict) -> dict:
             chat_template_kwargs = dict(payload_kwargs.get("chat_template_kwargs") or {})
             chat_template_kwargs.setdefault("enable_thinking", enable_thinking)
             payload_kwargs["chat_template_kwargs"] = chat_template_kwargs
+        payload_kwargs = _strip_falsy_logprobs(payload_kwargs)
         return payload_kwargs
 
     @with_circuit_breaker("llm")
@@ -200,6 +217,7 @@ async def generate(self, prompt: str, **kwargs) -> dict:
         base_url, model, headers = self._resolve_overrides(kwargs)
         kwargs.pop("metadata", None)
         payload = {**self._defaults, **kwargs, "model": model, "prompt": prompt}
+        payload = _strip_falsy_logprobs(payload)
         try:
             resp = await self._client.post(f"{base_url}/completions", json=payload, headers=headers)
             resp.raise_for_status()

tests/unit/services/inference/test_vllm_client.py

@@ -233,6 +233,63 @@ def capture(req: httpx.Request) -> httpx.Response:
 
         assert "max_retries" not in captured
 
+    @pytest.mark.asyncio
+    async def test_falsy_config_logprobs_not_forwarded_on_chat(self):
+        """The DI container forwards LLMParamsConfig.logprobs (default False)
+        into self._defaults. `logprobs: false` is the OpenAI default, so it adds
+        nothing — but strict providers whose schema lacks the field reject it by
+        name whatever the value, e.g. Gemini"""
+        captured: dict = {}
+
+        def capture(req: httpx.Request) -> httpx.Response:
+            captured.update(json.loads(req.content))
+            return _chat_response()
+
+        await self._make_client(capture, logprobs=False).chat([{"role": "user", "content": "hi"}])
+
+        assert "logprobs" not in captured
+
+    @pytest.mark.asyncio
+    async def test_falsy_config_logprobs_not_forwarded_on_generate(self):
+        captured: dict = {}
+
+        def capture(req: httpx.Request) -> httpx.Response:
+            captured.update(json.loads(req.content))
+            return _completions_response()
+
+        await self._make_client(capture, logprobs=False).generate("hi")
+
+        assert "logprobs" not in captured
+
+    @pytest.mark.asyncio
+    async def test_falsy_request_logprobs_dropped_with_top_logprobs(self):
+        """A client sending `logprobs: false` alongside `top_logprobs` must not
+        leak either: top_logprobs is only meaningful when logprobs is on."""
+        captured: dict = {}
+
+        def capture(req: httpx.Request) -> httpx.Response:
+            captured.update(json.loads(req.content))
+            return _chat_response()
+
+        await self._make_client(capture).chat([{"role": "user", "content": "hi"}], logprobs=False, top_logprobs=3)
+
+        assert "logprobs" not in captured
+        assert "top_logprobs" not in captured
+
+    @pytest.mark.asyncio
+    async def test_truthy_logprobs_forwarded(self):
+        """An explicit opt-in must keep flowing through unchanged."""
+        captured: dict = {}
+
+        def capture(req: httpx.Request) -> httpx.Response:
+            captured.update(json.loads(req.content))
+            return _chat_response()
+
+        await self._make_client(capture).chat([{"role": "user", "content": "hi"}], logprobs=True, top_logprobs=5)
+
+        assert captured["logprobs"] is True
+        assert captured["top_logprobs"] == 5
+
     @pytest.mark.asyncio
     async def test_trailing_slash_stripped(self):
         c = VLLMClient(endpoint="http://vllm:8000/v1/", model_name="m")