Skip to content

faiss_backend

diapason.tools.storage.faiss_backend

FAISS dense retrieval memory backend.

Uses cosine similarity via inner-product search on L2-normalised vectors. Requires faiss-cpu (or faiss-gpu) and numpy.

FAISSMemory

Bases: MemoryBackend

Dense retrieval backend powered by FAISS.

Stores document embeddings in a faiss.IndexFlatIP index (inner-product, which equals cosine similarity when vectors are L2-normalised before insertion/search).

clear()

Reset the index and all internal storage.

delete(doc_id)

Soft-delete doc_id. Return True if it existed.

retrieve(query, *, top_k=5, **kwargs)

Embed query and return the top-k most similar docs.

store(content, *, source='', metadata=None)

Embed and store content, returning a unique doc id.