FAISS-FACEBOOK

FAISS Facebook AI vector library backbone OSS.

Définition

FAISS index types selection : (1) IndexFlatL2 + IndexFlatIP : exact brute-force search, baseline accuracy 100% mais slow large datasets. (2) IndexIVFFlat : Inverted File partitioning (k-means clusters), faster than Flat, approximate, ~5-10x faster. (3) IndexIVFPQ : IVF + Product Quantization compression, 10-100x faster + 8-16x memory reduction with accuracy tradeoff. (4) IndexHNSWFlat : Hierarchical Navigable Small World graph, very fast + high accuracy, more memory. (5) IndexPQ : pure Product Quantization, less common. (6) IndexLSH : Locality Sensitive Hashing, special use cases. (7) IndexOPQ : Optimized Product Quantization, better accuracy. (8) IndexBinaryFlat / BinaryHNSW : binary vectors (Hamming distance). (9) IndexIVFPQR : refined IVFPQ with re-ranking. Distance metrics : L2, Inner Product, Cosine (via normalization). GPU : faiss-gpu Python package, CUDA acceleration up to 100x speedup vs CPU. Adoption : reference library all major vector DBs underlying, Meta production Instagram + WhatsApp + Facebook search recommendations.

Origine

FAISS initial release fevrier 2017 par Meta AI Research (Facebook Engineering Blog announcement) ; Herve Jegou + Matthijs Douze + Jeff Johnson + others authors ; ~30000+ GitHub stars 2024 ; foundational library ecosystem vector DBs.

Exemple en contexte

Instagram Reels recommendations uses FAISS internally : embed videos via custom Vision Transformer model, FAISS IVFPQ index 1B+ video embeddings, GPU-accelerated similarity search GPU clusters, returns recommended Reels candidates personalised ; latency <100ms ; serves 2B+ Instagram users.

Termes liés

• Milvus — uses FAISS underlying.