[faysou]

Summary

LiteLLM indexing_params and query_params now accept prompt, a literal prefix prepended to each text before the embedding request. This lets models that expect task prefixes, such as EmbeddingGemma (title: none | text: for documents, task: code retrieval | query: for queries), run behind OpenAI-compatible servers like llama.cpp, which apply no model prompts of their own.

embedding:
  provider: litellm
  model: openai/embeddinggemma-2
  indexing_params:
    prompt: 'title: none | text: '
  query_params:
    prompt: 'task: code retrieval | query: '
envs:
  OPENAI_BASE_URL: http://127.0.0.1:8080/v1

Design

prompt is applied in PacedLiteLLMEmbedder.run_embedding_request and never forwarded to litellm.aembedding. create_embedder passes indexing_params to the constructor as default kwargs, so handling it at the request boundary covers both paths: a per-call prompt (query side) overrides the constructor default (indexing side), and neither reaches the server as an unknown API parameter.

prompt joins input_type in the accepted LiteLLM keys, the ccc init template comment, and the README.

Validation

• uv run pytest tests/test_litellm_embedder.py tests/test_embedder_params.py tests/test_embed_params_forwarding.py tests/test_settings.py: 98 passed.
• uv run ruff check src tests and uv run mypy src pass.
• Indexed a large multi-repository workspace through llama.cpp serving EmbeddingGemma 2 with these prompts; ccc doctor passes both model checks.