FAISS

// Overview

FAISS (Facebook AI Similarity Search) is a library for efficient similarity search and clustering of dense vectors. It is written in C++ with full Python and numpy wrappers, and is developed primarily by Meta's Fundamental AI Research group. You give it a set of vectors, and it finds the ones closest to a query using L2 (Euclidean) distance or dot product, which also covers cosine similarity on normalized vectors.

It is built for developers working with embeddings, such as semantic search, recommendation, and retrieval-augmented generation. As an embedded, in-process library rather than a standalone server, you call it directly from your own code and keep the index in memory or on disk. This makes it a common building block underneath higher-level vector databases.

FAISS offers a range of index types so you can trade off search speed, accuracy, and memory. Exact indexes like IndexFlatL2 give precise results, while compressed and graph-based indexes (such as HNSW and NSG) scale to billions of vectors. Some of the algorithms also run on the GPU, where GPU indexes can act as drop-in replacements for their CPU counterparts.

// What it does

• Exact and approximate nearest-neighbor search over dense vectors using L2 distance or dot product
• Many index types that trade off search time, accuracy, and memory per vector
• Compressed and quantized indexes that scale to billions of vectors in RAM on a single server
• Graph-based indexes such as HNSW and NSG for faster approximate search
• Optional GPU implementation (CUDA or AMD ROCm) where GPU indexes drop in for CPU ones, with single and multi-GPU support
• C++ core with complete Python/numpy wrappers, plus k-means clustering and parameter-tuning utilities

// Getting started

FAISS ships precompiled for Anaconda. Install the CPU package, then build a flat index and search it. The README points to the project wiki for the full tutorial.

conda install -c pytorch faiss-cpu

import numpy as np
import faiss

d = 64                           # vector dimension
xb = np.random.random((10000, d)).astype('float32')  # database
xq = np.random.random((5, d)).astype('float32')      # queries

index = faiss.IndexFlatL2(d)     # exact L2 index
index.add(xb)                    # add vectors

D, I = index.search(xq, k=4)     # 4 nearest neighbors
print(I)                          # neighbor ids
print(D)                          # squared L2 distances

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Powering semantic search over embeddings, where you find the documents closest to a query vector
• Serving as the in-process retrieval layer for a retrieval-augmented generation (RAG) pipeline
• Running nearest-neighbor search at scale, up to billions of vectors, using compressed or GPU indexes
• Clustering dense vectors with the built-in k-means implementation

// How FAISS compares

FAISS alongside other open-source vector databases tools AI/TLDR tracks, ranked by GitHub stars.