You could first calculate the distance of the first n bits (eg: 64, one popcountll) as a first pass, then calculate the full distance for candidates over a threshold from the first pass. It makes it approximate, but depending on the application it can be worth it.
mbreese 3 hours ago [-]
I was thinking of something similar — instead of just two passes, couldn’t you also store different quantized values? If you have thousands of documents, you could narrow it down to a handful with a few bit-wise Hamming comparisons before doing a full cosine similarity on just the rest. If you hand more than one bitmap stored, you’d have fewer comparisons at each step too.
Author here... it is free for open-source projects.
What kind of license would you like more?
stevesimmons 3 hours ago [-]
USearch has a sqlite extension that supports various metrics on including Hamming distance on standard sqlite BLOB columns. It gets similar performance and is very convenient.
(There's also an indexed variant that does faster lookups, but it uses a special virtual table layout that constrains the types of the other columns in the table.)
Has anyone tried keyword expansion in this context?
I had the idea of making a "poor man's embeddings" for document similarity. You want two documents to match even if they share no keywords, as long as their keywords are closely related, right? That seems like a very solvable problem.
esafak 2 hours ago [-]
Are today's models any good at helping write postgres or sqlite extensions?
Rendered at 19:06:28 GMT+0000 (Coordinated Universal Time) with Vercel.
Would this work?
What kind of license would you like more?
(There's also an indexed variant that does faster lookups, but it uses a special virtual table layout that constrains the types of the other columns in the table.)
See https://github.com/unum-cloud/USearch. pip-installable for Python users.
I had the idea of making a "poor man's embeddings" for document similarity. You want two documents to match even if they share no keywords, as long as their keywords are closely related, right? That seems like a very solvable problem.