Using an embedding model from Hugging Face

[Nikita Toropov (tossik8)]

I get the following error every time I attempt to use an embedding model from Hugging Face:

Failed to validate embedding model (default_embedding_model) params litellm.BadRequestError: HuggingfaceException - {"error":"Model not supported by provider hf-inference"}

settings.yaml:

embedding_models:
  default_embedding_model:
    model_provider: huggingface 
    model: BAAI/bge-small-en-v1.5
    auth_method: api_key
    api_key: ${HUGGINGFACE_API_KEY}

I have tried:

• adding huggingface/ to model
• removing BAAI/ from model
• adding /BAAI to model_provider
• changing the model itself

Nothing seems to help. Has anyone faced this issue before? How did you resolve it?

[Juyel (juyelhushen)]

I think the issue is with the Hugging Face Inference Provider rather than the BAAI/bge-small-en-v1.5 model name itself.

BAAI/bge-small-en-v1.5 is a valid embedding model, but Hugging Face's hf-inference provider may not currently support that model for the embedding request you're making. That's why changing the model string alone doesn't resolve the Model not supported by provider hf-inference error.

I'd first verify whether the model has an available inference provider for the feature-extraction/embedding task in Hugging Face. If hf-inference isn't available, you may need to select another supported provider or use a dedicated Hugging Face Inference Endpoint.

Also, I would keep the GraphRAG configuration conceptually as:

embedding_models:
  default_embedding_model:
    model_provider: huggingface
    model: BAAI/bge-small-en-v1.5
    auth_method: api_key
    api_key: ${HUGGINGFACE_API_KEY}

rather than adding huggingface/ to the model name. GraphRAG's configuration treats model_provider as the provider and model as the model identifier.

If you can share your GraphRAG and LiteLLM versions, I can check the exact Hugging Face provider configuration for those versions.

[Nima Simyari (NimaSimyari)]

I don't think the model identifier itself is the problem here. Your GraphRAG configuration is structurally correct: model_provider should contain the LiteLLM provider (huggingface), while model should contain the Hugging Face model ID (BAAI/bge-small-en-v1.5).

The error

Model not supported by provider hf-inference

is coming from the Hugging Face inference layer rather than from GraphRAG's YAML parsing.

BAAI/bge-small-en-v1.5 is currently available for feature extraction through the Hugging Face hf-inference provider, so I would first try upgrading GraphRAG/LiteLLM and keep the configuration simple:

embedding_models:
  default_embedding_model:
    model_provider: huggingface
    model: BAAI/bge-small-en-v1.5
    auth_method: api_key
    api_key: ${HUGGINGFACE_API_KEY}

I would not add huggingface/ to the model field, because GraphRAG already separates the provider from the model identifier.

If the same error persists, I would test the model directly against Hugging Face first. That helps distinguish a GraphRAG/LiteLLM integration issue from an HF provider-availability issue.

Another reliable option is to deploy the model as a dedicated Hugging Face Inference Endpoint and provide its URL through api_base, e.g.:

embedding_models:
  default_embedding_model:
    model_provider: huggingface
    model: BAAI/bge-small-en-v1.5
    auth_method: api_key
    api_key: ${HUGGINGFACE_API_KEY}
    api_base: https://<your-endpoint>.huggingface.cloud

So in short: I would keep the original provider/model split, update the GraphRAG/LiteLLM stack, and verify the HF inference route rather than changing the BGE model name.