For this patch we no longer manually do the bit quantization and instead rely on built in sqlite-vec features. Running the performance tests with manual rowLimit set to 10,000 (a typical use case) on my MacM1 semantic searching latency was reduced 3x, from 6.5 to 2.2ms. I have added one additional performance test that uses multiple different queries, because I thought one query was insufficient. In the original we were searching 100 entries for 2 results. That is a '50x' oversample, which is being set manually each time we do a query with a special function. Note there is also a 're-vendor' to upgrade sqlite-vec because they fixed a bug we found in the integration of this patch. Base revision's try run: https://treeherder.mozilla.org/jobs?repo=try&landoCommitID=193651 Local revision's try run: https://treeherder.mozilla.org/jobs?repo=try&landoCommitID=194008 Differential Revision: https://phabricator.services.mozilla.com/D300160
2 lines
15 B
Plaintext
2 lines
15 B
Plaintext
0.1.10-alpha.4
|