Hacker News Reader

I've recently presented the approach with exhaustive search and heap, then iteratively adding simple indices, using ClickHouse. It works pretty well despite the overall usage simplicity.

Video: https://www.youtube.com/watch?v=hGRNcftpqAk

Presentation: https://presentations.clickhouse.com/meetup74/ai/

PS. It should work at the same speed as "bruteforce-blas" in the ANN benchmarks.

View on HN

nice, if I'm reading this right, you do random projection as a first stage filter, then exact search over the filtered subset? that's clever, did you test precision/recall at all?

Yes, this is the idea. I set the threshold for approximate search with a margin so that on a few test queries, I get 100% recall, and then it is filtered by precise search.

But I didn't evaluate it thoroughly. The threshold for Hamming distance over bit masks was set almost arbitrarily, and it will be interesting to test further.

Reply on news.ycombinator.com