Hacker News Reader

I'd be interested to understand the query performance when compared to the HNSW implementation (Turso?) they mentioned. In general search performance is more important to me, and I don't mind having an increase in insert overhead to have very fast vector search.

View on HN

HNSW is not accurate. I guess brute-force means that sqlite-vector returns the best match.

Right but libsql(Turso) uses HNSW - so i'd be curious to know how performance of sqlite-vector compares - they do say they "Use a brute-force-like approach, but highly optimized." - which to me, would be very interesting to see compared with a HNSW approach.

I believe a common approach with this kind of inaccurate index is to use the index to get the top 100 and then calculate the exact distance against those 100 matches to get the top 10.

Reply on news.ycombinator.com