Hello, I use Solr 9.8 vector search to find fragments of text to which I associate vectors.
To do this, I use the “knn Query Parser”. Until now, I used a “Top-K” of 1000. The ratio between performance and relevance seemed perfect to me. Now, I have just over a million vectors and the latency is becoming prohibitive (over 10 seconds). So I've started lowering the “Top-K” (since I don't really need 1,000 results). Although response times are greatly improved by lowering the “Top-K”, I'm finding that responses are less relevant: Very good answers are no longer returned with a lower “Top-K”. I assume that this type of “tradeoff” is intrinsically linked to the very nature of the HNSW algorithm. However, I'd like to test the role of the “hnswMaxConnections” and “hnswBeamWidth” parameters to see if they can help me find a good relevance/performance balance. Are these parameters involved in indexing? During search? At both stages? Just to find out whether it's necessary to re-index all the vectors between each test. If not, what factors would be most significant in improving performance? It seems to me that the Solr core doesn't consume much memory. Perhaps it's possible to allocate more memory to the HNSW algorithm (or to the caches linked to the vectors)? Thanks for your attention, Guillaume
