Author here, happy to answer any questions! Been working on this for a while, so I'm very happy to get this v0.1.0 "stable" release out.
sqlite-vec works on MacOS, Linux, Windows, Raspberry Pis, in the browser with WASM, and (theoretically) on mobile devices. I focused a lot on making it as portable as possible. It's also pretty fast - benchmarks are hard to do accurately, but I'd comfortable saying that it's a very very fast brute-force vector search solution.
One experimental feature I'm working on: You can directly query vectors that are in-memory as a contiguous block of memory (ie NumPy), without any copying or cloning. You can see the benchmarks for that feature here under "sqlite-vec static", and it's competitive with faiss/usearch/duckdb https://alexgarcia.xyz/blog/2024/sqlite-vec-stable-release/i...
On the releases page https://github.com/asg017/sqlite-vec/releases/tag/v0.1.0 can you explain what is vec0.dll vs sqlite-vec-0.1.0-loadable-windows-x86_64.tar.gz, which also contains a similarly named vec0.dll but of a different size?
will fix up those links this weekend, thanks for pointing it out!
the cvec0.all file can be ignored. i was debugging a issue and uploaded a mscv compiled version as that, but its not part of the official release. prefer the tar.gz version and let me know if it doesn't work
Can't seem to make the loadable version work on Windows client (using .load and the latest x64 version of sqlite).
On Windows 10 it says "Error: the specified module could not be found."
On Windows 7 the error is more detailed and complains about missing libgcc_s_seh-1.dll which, according to a quick Google search, may or may not result from a compilation problem with MinGW...?
Great to see this. Seems simple enough, but I can't wait until ORMs like peewee incorporate support alongside things like FTS, etc. just for the sake of case of use.
I feel like I've touched a lot of things where something like this is useful (hobby projects). In my case I've done a recommendation engine, music matching (I specifically use it for matching anime to their data), and perceptual hash matching.
Really curious to hear about what kind of music embedding models/tools you used! I've tried finding some good models before but they were all pretty difficult to use
In the past I mainly screwed around with [2] musig as it was fairly easy to modify, right off hand I don't know why I stopped using it. [1] seek-tune (not-shazam) was just post here on hn yesterday, I had to modify it quite a bit to actually get something working in a way I was happy with. Overall it was a little slow, but that's not something I'm gonna fault the author for as the project is still quite new. While browsing around yesterday after messing with [2], I found [3] cpuimage/shazam. It's a little easier for me to screw around with as it's in c, and would probably be the one I try and modify to use this project.
Everything is based on the shazam algorithm or something similar, so chunking the audio stream, and fingerprinting it. I was focused on that because it allows me to figure out where OP and ED are in some anime and allows me to add a skip button. Also since anime in the same season typically have the same OP/ED I can use that to try and auto categorize them after the first ones been identified.
TL;DR: It's all shazam look alikes, everything else was to annoying to use.
Not really sure what I can give you. The tools were basically just golang, gcc and docker. Everything else inspection wise just came while I was writing the code. The viewer I use is jellyfin, but I didn't get far enough in the project to have a functioning plugin for it. If you want more info feel free to email me or shoot me a friend request on discord (both visible on my profile).
I absolutely love this, great work! For those that might find it useful, I created a Python notebook that shows how to extend this to perform Hybrid Search (Vector + BM25 based Full Text search) https://github.com/liamca/sqlite-hybrid-search
Duckdb is an excellent choice for this task, and it’s incredibly fast!
We’ve also added vector search to our product, which is really useful.
OpenAI’s official examples of embedding search use cosine similarity. But here’s the cool part: since OpenAI embeddings are unit vectors, you can just run the dot product instead!
DuckDB has a super fast dot product function that you can use with SQL.
In our product, we use duckdb-wasm to do vector searches on the client side.
What library do you use to compute embedding right now? I'm wondering if it's possible to do both embedding and vector search client-side for a local-first app.
You can generate "text embeddings" (ie vectors) from your text with an embedding model. Once you have your text represented as vectors, then "closest vector" means "more semantically similar", as defined by the embedding model you use. This can allow you to build semantic search engines, recommendations, and classifiers on top of your text and embeddings. It's kindof like a fancier and fuzzier keyword search.
I wouldn't completely replace keyword search with vector search, but it's a great addition that essentially lets you perform calculations on text
Nice explanation. One use case where keywords haven't worked well for me , and (at least at first glance) vectors are doing better are longer passages -- finding sentences that are similar rather than just words.
Do embeddings work across languages? It must depend on the model, but are there models where similar concepts in different languages occupy the same location in vector space?
If yes it would allow to search a corpus in one language using another, without translating anything up front (neither the corpus nor the query).
One use case is for retrieving similar documents, such as when recommending related content. Another use case is retrieving document segments that are similar to a user query and passing them along with the user query to a large language model for improvement in the generated response. Vector search is also better in some ways than keyword search since it can find documents that are semantically similar even when the user may not have used the exact keyword, or even partial keywords like “Postgres” instead of “PostgreSQL.”
I love this. I know how much work addressing the dependencies must be, but you’re really attacking the right problems. Looking forward to trying this out with my project.
No, libsql added custom vector search directly into their library, while sqlite-vec is a separate SQLite extension.
The libsql vector feature only works in libsql, sqlite-vec works in all SQLite versions. The libsql vector feature works kindof like pgvector, while sqlite-vec works more like the FTS5 full text SQLite extension.
I'd say try both and see which one you like more. sqlite-vec will soon be a part of Turso's and SQLite Cloud's products.
vec0 virtual tables have a hard-coded max of 8192 dimensions, but I can raise that very easily (I wanted to reduce resource exhaustion attacks). But if you're comparing vectors manually, then the `vec_distance_ls()` and related functions have no limits (besides SQLite's 1GB blob limit)
sqlite-vec works on MacOS, Linux, Windows, Raspberry Pis, in the browser with WASM, and (theoretically) on mobile devices. I focused a lot on making it as portable as possible. It's also pretty fast - benchmarks are hard to do accurately, but I'd comfortable saying that it's a very very fast brute-force vector search solution.
One experimental feature I'm working on: You can directly query vectors that are in-memory as a contiguous block of memory (ie NumPy), without any copying or cloning. You can see the benchmarks for that feature here under "sqlite-vec static", and it's competitive with faiss/usearch/duckdb https://alexgarcia.xyz/blog/2024/sqlite-vec-stable-release/i...