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

[genisis0x]

Summary

• Closes [BUG] RD-Agent/LiteLLM: "LLM Provider NOT provided" even with correct provider/model for Ollama/DeepSeek (env, patch, CLI all fail) #1016. When `LITELLM_CHAT_MODEL` is set to a bare name LiteLLM cannot map to a known provider (`deepseek-chat`, `qwen2:7b`, ...), the eventual `completion()` call raises `BadRequestError: LLM Provider NOT provided` ten retries deep inside the base `APIBackend` retry loop, with no hint that the fix is to prefix the model with `/`. The reporter (and four other users in the thread) saw this as "impossible to use Ollama / DeepSeek / SiliconFlow even though docs say it is supported".

Fix

• Validate the chat and embedding model strings up front in `LiteLLMAPIBackend.init` by calling `litellm.utils.get_llm_provider`.
• If it raises `BadRequestError`, reraise as `RuntimeError` with a clear message that names the offending env var (`CHAT_MODEL` / `EMBEDDING_MODEL`), shows the bad value, and lists corrected examples (`openai/gpt-4o`, `deepseek/deepseek-chat`, `ollama/qwen2:7b`, `ollama/nomic-embed-text`). The original `BadRequestError` is preserved as `cause` for advanced debugging.

Compatibility

• The default `chat_model="gpt-4-turbo"` and `embedding_model="text-embedding-3-small"` continue to resolve through LiteLLM's bare-name fallback and remain valid.
• Any user who already has `CHAT_MODEL=ollama/qwen2:7b` etc. is unaffected.
• Users who previously hit the buried `LLM Provider NOT provided` error will now see a clear actionable message before any HTTP call.

Tests

• New `test/oai/test_litellm_model_validation.py` covers: default chat / embedding names pass; provider-prefixed names (openai, deepseek, ollama, azure) pass; bare unknown names raise `RuntimeError` whose message mentions the env var, the bad value, and the `/` hint; original `BadRequestError` is preserved as `cause`.
• No live LLM endpoint required — tests use the same client-side `get_llm_provider` resolution the validation function uses.

Checklist

• Issue linked
• Tests added
• No behavioural change for the default model configuration
• CLA will be confirmed in a follow-up comment

---

📚 Documentation preview 📚: https://RDAgent--1412.org.readthedocs.build/en/1412/

[genisis0x]

Read the CLA — all clear from my side.

@microsoft-github-policy-service agree