fix(oai): surface actionable error when CHAT_MODEL/EMBEDDING_MODEL lack a LiteLLM provider prefix

rdagent/oai/backend/litellm.py

@@ -12,6 +12,7 @@
     token_counter,
 )
 from litellm.exceptions import BadRequestError, Timeout
+from litellm.utils import get_llm_provider
 from pydantic import BaseModel
 
 from rdagent.log import LogColors
@@ -45,12 +46,44 @@ class Config:
 ACC_COST = 0.0
 
 
+def _validate_litellm_model_provider(model: str, setting_env_var: str) -> None:
+    """Validate that ``model`` resolves to a LiteLLM provider before any
+    completion / embedding call is attempted.
+
+    LiteLLM uses a ``<provider>/<model>`` naming convention; for bare names it
+    cannot disambiguate it falls back to the OpenAI provider, and for names
+    it does not recognise (e.g. ``deepseek-chat``, ``qwen2:7b``) it raises
+    ``BadRequestError: LLM Provider NOT provided``. That error surfaces from
+    deep inside the retry loop with no actionable hint, which is the bug
+    reported in #1016. By calling ``get_llm_provider`` at backend
+    construction, the misconfiguration is caught before any HTTP call with a
+    clear message that points at the env-var fix.
+    """
+    try:
+        get_llm_provider(model)
+    except BadRequestError as e:
+        raise RuntimeError(
+            f"Could not determine the LiteLLM provider for "
+            f"{setting_env_var}={model!r}.\n"
+            f"LiteLLM uses a '<provider>/<model>' naming convention. "
+            f"Update the model name to include the provider prefix, for example:\n"
+            f"  CHAT_MODEL=openai/gpt-4o\n"
+            f"  CHAT_MODEL=deepseek/deepseek-chat\n"
+            f"  CHAT_MODEL=ollama/qwen2:7b\n"
+            f"  EMBEDDING_MODEL=ollama/nomic-embed-text\n"
+            f"See https://docs.litellm.ai/docs/providers for the full list "
+            f"of supported provider prefixes."
+        ) from e
+
+
 class LiteLLMAPIBackend(APIBackend):
     """LiteLLM implementation of APIBackend interface"""
 
     _has_logged_settings: bool = False
 
     def __init__(self, *args: Any, **kwargs: Any) -> None:
+        _validate_litellm_model_provider(LITELLM_SETTINGS.chat_model, "CHAT_MODEL")
+        _validate_litellm_model_provider(LITELLM_SETTINGS.embedding_model, "EMBEDDING_MODEL")
         if not self.__class__._has_logged_settings:
             logger.info(f"{LITELLM_SETTINGS}")
             logger.log_object(LITELLM_SETTINGS.model_dump(), tag="LITELLM_SETTINGS")

test/oai/test_litellm_model_validation.py

@@ -0,0 +1,65 @@
+"""Unit tests for the upfront ``<provider>/<model>`` validation added in
+response to #1016.
+
+These tests do not hit any live LLM endpoint; they only exercise the
+``get_llm_provider`` resolution that LiteLLM exposes purely client-side.
+"""
+
+import unittest
+
+import pytest
+
+# get_llm_provider is the same helper the validation function uses; this
+# avoids tripping over the lazy environment requirements of constructing a
+# full LiteLLMAPIBackend instance.
+try:
+    from litellm.exceptions import BadRequestError
+    from litellm.utils import get_llm_provider
+except ImportError:  # pragma: no cover - litellm is a hard dependency of RD-Agent
+    pytest.skip("litellm is not installed; skipping provider validation tests", allow_module_level=True)
+
+from rdagent.oai.backend.litellm import _validate_litellm_model_provider
+
+
+class TestValidateLiteLLMModelProvider(unittest.TestCase):
+    def test_default_chat_model_passes(self) -> None:
+        # gpt-4-turbo is the default; LiteLLM resolves bare OpenAI names.
+        _validate_litellm_model_provider("gpt-4-turbo", "CHAT_MODEL")
+
+    def test_default_embedding_model_passes(self) -> None:
+        _validate_litellm_model_provider("text-embedding-3-small", "EMBEDDING_MODEL")
+
+    def test_provider_prefixed_chat_models_pass(self) -> None:
+        for model in [
+            "openai/gpt-4o",
+            "deepseek/deepseek-chat",
+            "ollama/qwen2:7b",
+            "azure/my-deployment",
+        ]:
+            with self.subTest(model=model):
+                _validate_litellm_model_provider(model, "CHAT_MODEL")
+
+    def test_bare_unknown_chat_model_raises_actionable_error(self) -> None:
+        # This is the exact failure mode from #1016: a bare model name that
+        # LiteLLM cannot map to a known provider.
+        with pytest.raises(RuntimeError) as exc_info:
+            _validate_litellm_model_provider("definitely-not-a-real-model-xyz", "CHAT_MODEL")
+        message = str(exc_info.value)
+        # Error must surface the offending env var name, the bad value, and
+        # at least one corrected example, so the user can fix it without
+        # having to grep the source.
+        self.assertIn("CHAT_MODEL", message)
+        self.assertIn("definitely-not-a-real-model-xyz", message)
+        self.assertIn("<provider>/<model>", message)
+        # The chained cause should still be the original LiteLLM exception so
+        # advanced users can inspect it.
+        self.assertIsInstance(exc_info.value.__cause__, BadRequestError)
+
+    def test_bare_unknown_embedding_model_raises_actionable_error(self) -> None:
+        with pytest.raises(RuntimeError) as exc_info:
+            _validate_litellm_model_provider("definitely-not-a-real-emb-xyz", "EMBEDDING_MODEL")
+        self.assertIn("EMBEDDING_MODEL", str(exc_info.value))
+
+
+if __name__ == "__main__":
+    unittest.main()