I’m writing a new vector search SQLite Extension (alexgarcia.xyz)

Article summary

The author is developing a new SQLite extension called sqlite-vec for vector search, which will replace the existing sqlite-vss extension. The new extension is written in C, has no dependencies, and will provide custom SQL functions and virtual tables for fast vector search. It will support various features such as quantization, JSON/BLOB/numpy conversions, and vector arithmetic. The extension is still in development, but the author plans to release the first version soon.

Main themes

• Vector Search
• SQLite Extension
• Database Optimization
• WASM Support
• Quantization
• Performance Comparison

What commenters say

• The author's approach to developing the sqlite-vec extension is seen as a good example of 'readme driven development'.
• Some commenters are interested in the potential for the extension to support indexing strategies like HNSW and IVF.
• There is a discussion about the trade-offs between storage size and search performance, with some commenters suggesting that the extension could benefit from compressed vector storage.
• The author's decision to write the extension in C rather than Rust is debated, with some commenters arguing that Rust would have been a better choice.
• Some commenters are excited about the potential for the extension to be used in various applications, such as local AI apps and vector databases.
• There is a comparison between the performance of the sqlite-vec extension and other vector search libraries, such as Faiss and HNSWlib.
• The author's plan to add support for approximate nearest neighbors (ANN) options in the future is seen as a positive development.
• Some commenters are interested in the potential for the extension to be used with other databases, such as Postgres, and for it to support a common SQL vector DSL.