The Faiss library is a powerful toolkit for vector similarity search, designed to efficiently handle large collections of embedding vectors. It provides indexing methods and related primitives for searching, clustering, compressing, and transforming vectors. The paper discusses the tradeoff space of vector search, the design principles of Faiss, and benchmarks key features of the library. It also highlights applications of Faiss in various domains, including industrial settings, text retrieval, data mining, and content moderation. Faiss is an industrial-grade library for approximate nearest neighbor (ANNS) search. It is designed to be used both from simple scripts and as a building block of a database management system (DBMS). Unlike other libraries that focus on a single indexing method, Faiss is a toolbox that contains indexing methods that commonly involve a chain of components, such as preprocessing, compression, and non-exhaustive search. This flexibility is necessary because the most efficient indexing methods vary depending on the usage constraints. Faiss is not a feature extractor, a service, or a database. It only indexes embeddings that have been extracted by a different mechanism. The library's scope is intentionally limited to focus on carefully implemented algorithms. The basic structure of Faiss is the index, which can store a number of database vectors that are progressively added to it. At search time, a query vector is submitted to the index, and the index returns the database vector that is closest to the query vector in terms of Euclidean distance. The paper discusses the performance axes of a vector search library, including accuracy, memory usage, and speed. It also explores the tradeoffs between these axes and the different indexing methods available in Faiss. The paper highlights the importance of vector compression and non-exhaustive search in achieving efficient and accurate results. It also discusses the various types of quantizers used in Faiss, including scalar quantizers, product quantizers, and additive quantizers, and their respective tradeoffs in terms of accuracy and code size. The paper also discusses the different types of non-exhaustive search methods available in Faiss, including inverted file and graph-based methods. These methods are designed to efficiently search large datasets by focusing on a subset of database vectors that are most likely to contain the search results. The paper also discusses the importance of preprocessing steps, such as principal component analysis (PCA) and orthogonal projection, in improving the performance of Faiss. In conclusion, Faiss is a powerful library for vector similarity search that provides a wide range of indexing methods and related primitives. It is designed to be flexible and efficient, allowing users to balance accuracy, memory usage, and speed based on their specific needs. The library is widely used in various applications, including industrial settings, text retrieval, data mining, and content moderation.The Faiss library is a powerful toolkit for vector similarity search, designed to efficiently handle large collections of embedding vectors. It provides indexing methods and related primitives for searching, clustering, compressing, and transforming vectors. The paper discusses the tradeoff space of vector search, the design principles of Faiss, and benchmarks key features of the library. It also highlights applications of Faiss in various domains, including industrial settings, text retrieval, data mining, and content moderation. Faiss is an industrial-grade library for approximate nearest neighbor (ANNS) search. It is designed to be used both from simple scripts and as a building block of a database management system (DBMS). Unlike other libraries that focus on a single indexing method, Faiss is a toolbox that contains indexing methods that commonly involve a chain of components, such as preprocessing, compression, and non-exhaustive search. This flexibility is necessary because the most efficient indexing methods vary depending on the usage constraints. Faiss is not a feature extractor, a service, or a database. It only indexes embeddings that have been extracted by a different mechanism. The library's scope is intentionally limited to focus on carefully implemented algorithms. The basic structure of Faiss is the index, which can store a number of database vectors that are progressively added to it. At search time, a query vector is submitted to the index, and the index returns the database vector that is closest to the query vector in terms of Euclidean distance. The paper discusses the performance axes of a vector search library, including accuracy, memory usage, and speed. It also explores the tradeoffs between these axes and the different indexing methods available in Faiss. The paper highlights the importance of vector compression and non-exhaustive search in achieving efficient and accurate results. It also discusses the various types of quantizers used in Faiss, including scalar quantizers, product quantizers, and additive quantizers, and their respective tradeoffs in terms of accuracy and code size. The paper also discusses the different types of non-exhaustive search methods available in Faiss, including inverted file and graph-based methods. These methods are designed to efficiently search large datasets by focusing on a subset of database vectors that are most likely to contain the search results. The paper also discusses the importance of preprocessing steps, such as principal component analysis (PCA) and orthogonal projection, in improving the performance of Faiss. In conclusion, Faiss is a powerful library for vector similarity search that provides a wide range of indexing methods and related primitives. It is designed to be flexible and efficient, allowing users to balance accuracy, memory usage, and speed based on their specific needs. The library is widely used in various applications, including industrial settings, text retrieval, data mining, and content moderation.