sqlite-vector
A production-grade vector extension for SQLite targeting enterprise scenarios. It provides more thorough performance optimizations and a rich set of tuning options, with support for quantization and multiple distance metric configurations.
Timeline (last 90 days)
When installing sqlite-vector via pip/uv, the package name is sqlite-ai-vector rather than sqlite-vector, though the extension itself is still called vector; installing the wrong package will result in a "no module named sqlite_vector" error.
sqlite-vec stores vectors using virtual tables, whereas sqlite-vector uses ordinary tables and stores vectors as BLOBs.
sqlite-vector requires an explicit one-step initialization of vector_init, specifying the table, column, data type, dimensions, and distance metric, and the distance metric supports cosine, Manhattan, Euclidean, and others.
In benchmarks, at a scale of 50,000 vectors, sqlite-vec takes approximately 84ms/query and sqlite-vector approximately 78ms/query; at 150,000 vectors, the figures are 241ms and 231ms respectively, with sqlite-vector generally querying faster.
In memory efficiency tests with 50,000 vectors, sqlite-vec requires approximately 150MB for full vector search, while sqlite-vector uses quantization (turbo4) and requires only approximately 19.8MB.
All Facts (5)
When installing sqlite-vector via pip/uv, the package name is sqlite-ai-vector rather than sqlite-vector, though the extension itself is still called vector; installing the wrong package will result in a "no module named sqlite_vector" error.
50%Unverifiedsqlite-vec stores vectors using virtual tables, whereas sqlite-vector uses ordinary tables and stores vectors as BLOBs.
50%Unverifiedsqlite-vector requires an explicit one-step initialization of vector_init, specifying the table, column, data type, dimensions, and distance metric, and the distance metric supports cosine, Manhattan, Euclidean, and others.
50%UnverifiedIn benchmarks, at a scale of 50,000 vectors, sqlite-vec takes approximately 84ms/query and sqlite-vector approximately 78ms/query; at 150,000 vectors, the figures are 241ms and 231ms respectively, with sqlite-vector generally querying faster.
50%UnverifiedIn memory efficiency tests with 50,000 vectors, sqlite-vec requires approximately 150MB for full vector search, while sqlite-vector uses quantization (turbo4) and requires only approximately 19.8MB.
50%