The built-in docs_folder and chunks_file retrievers use BM25 and need no embedding model. The project also has lazy-loaded adapters for Chroma, FAISS, Qdrant, pgvector, Pinecone, Weaviate, Milvus, Elastic/OpenSearch, and LanceDB.
Install an adapter only when you use it:
uv pip install --python .venv/bin/python -e '.[qdrant]'
# or: .[chroma], .[faiss], .[pgvector], .[pinecone], .[weaviate],
# .[milvus], .[elastic], .[lancedb]Vector-backed retrievers need an embeddings: section configured with the model used to build the index. init performs a retriever health check and sample search before it saves the project.
Anthropic is the default judge provider. OpenAI and OpenAI-compatible endpoints are supported through lazy.yaml or setup flags; the latter requires both a model name and base_url. Anthropic Batch API is used for Truth Check and Agent Trajectory when the corpus meets batch_min_examples; other tasks run synchronously.