abstracts-faiss

This is a faiss index, trained on abstracts-embeddings. A ready-to-go search interface for using this index is available at abstracts-index.

Building

It was trained with the train.py script found at abstracts-search with the options -N -c 65536 (normalized, train 65536 clusters), using the default preprocess technique OPQ96_384 (PCA to a 384-dimensional vector, then apply OPQ for a 96-byte code). Note that, although the Stella model was trained with Matryoshka (MRL) loss, it outputs ordinary vectors which are not expected to be truncated, so PCA was used.

Tuning

The index comes with the Pareto-optimal parameters from faiss.ParameterSpace.explore at index/params.json, so a point on the speed-recall tradeoff can be immediately picked. For reference, the exec_time field is seconds-per-query on an i7-12700H, and the recall field is 1-Recall@1. (1-Recall@1 is the probability that the top result returned by the index is the true closest, measured using a holdout set.) The best recall is 0.756 with a search time of 0.11 seconds, but a recall of 0.723 can be had with a search time of 0.0042 seconds.