FAISS

FAISS (Facebook AI Similarity Search) is a library developed by Facebook AI Research that enables efficient similarity search and clustering of dense vectors. It is especially well-suited for applications involving high-dimensional vector data, such as NLP

Related terms:

• Vector Embedding

Overview

FAISS is optimized for:

• Fast retrieval from large collections of vectors (millions or more).
• Approximate nearest neighbor (ANN) search, which trades off accuracy for speed.
• Exact search, depending on the chosen index type.
• GPU acceleration for very large-scale search tasks.

Core Concept

At its core, FAISS takes a large number of high-dimensional vectors (e.g., sentence or document embeddings), and enables fast similarity search to retrieve the most similar vectors to a given query.

For example, in an NLP context:

• Documents or notes are embedded into vector space using a model like SBERT.
• These embeddings are stored in a FAISS index.
• Given a query, its embedding is computed, and FAISS returns the nearest neighbors (i.e., most semantically similar notes).

Index Types

FAISS offers different types of indices depending on use case:

• IndexFlatL2: exact search using L2 (Euclidean) distance.
• IndexIVFFlat: approximate search using inverted files.
• IndexHNSW: Hierarchical Navigable Small World graph-based index (good for high recall).
• IndexPQ: product quantization for memory-efficient indexing.