Bug in LLMConfig and OpenAIEmbedderConfig instantiation: default models always loaded

[adamkatav]

Hello 👋 I'm working on adding graphiti docker compose MCP to my (pydantic AI based) agent with ollama local hosted models, mistral-small3.1 and snowflake-arctic-embed2. I passed the env vars to compose like so:

    environment:
      - OPENAI_API_KEY=None
      - OPENAI_BASE_URL=http://host.docker.internal:11434/v1
      - MODEL_NAME=mistral-small3.1
      - EMBEDDER_MODEL_NAME=snowflake-arctic-embed2
      - ... (Extra variables)

However the gpt-4.1-nano and text-embedding-3-small was chosen instead which caused this error: Error code: 404 - {'error': {'message': 'model "gpt-4.1-nano" not found, try pulling it first', 'type': 'api_error', 'param': None, 'code': None}}.

So I did some digging and found that:

• In graphiti_mcp_server.py at 304

llm_client_config = LLMConfig(api_key=self.api_key, model=self.model)

should be

llm_client_config = LLMConfig(api_key=self.api_key, model=self.model, small_model=self.model)

Because if small_model is not defined is defaults to gpt-4.1-nano

• On line 411

embedder_config = OpenAIEmbedderConfig(api_key=self.api_key, model=self.model)

Should be

embedder_config = OpenAIEmbedderConfig(api_key=self.api_key, embedding_model=self.model)

Because OpenAIEmbedderConfig has no model member, only embedding_model

After both patches, it works. Thanks :)

[prasmussen15]

Thanks for pointing out this issue in the MCP server! Would you like to open a PR with these as fixes? Otherwise I will be happy to fix them

[adamkatav]

@prasmussen15 Sure