Interestingly, while we don't fine-tune generative models for Louie.ai, we found fine-tuning embedding models to be a major $ saver. Instead of 1K-2K wide frontier embedding vector lens... Just 64. Huge savings on vector DB $$$.<p>I'm curious how that works with something like turboquant. Not needed any more, still dominant, better together, ... .
FAISS is no longer close to SoTA:<p><a href="https://ann-benchmarks.com/index.html" rel="nofollow">https://ann-benchmarks.com/index.html</a>
<a href="https://vector-index-bench.github.io/" rel="nofollow">https://vector-index-bench.github.io/</a>
<a href="https://big-ann-benchmarks.com/neurips23.html" rel="nofollow">https://big-ann-benchmarks.com/neurips23.html</a>
I think their point is the size/performance tradeoff rather than outright performance. The point of TurboQuant is the size savings, while still giving high accuracy.<p>It's been a while, but I do recall some high-performing vector matching indexes being very large.
Surprised that usearch isn't in any of these, it's pretty fast.
Wow! 4GB for 10 million documents. This means one could build a reverse index much faster than before and devx processes like debugging, performance testing would become much smoother. Can't wait for the sqlite bindings to come out!
It would be nice to have the README be a little more human written for a project where you actually want people to adopt it
people should read turboquant's open review comments: <a href="https://openreview.net/forum?id=tO3ASKZlok" rel="nofollow">https://openreview.net/forum?id=tO3ASKZlok</a>
tl,dr: there is an allegedly better alternative, and it's already implemented everywhere: <a href="https://github.com/VectorDB-NTU/RaBitQ-Library#rabitq-in-industry" rel="nofollow">https://github.com/VectorDB-NTU/RaBitQ-Library#rabitq-in-ind...</a>
This looks perfect for local, privacy first search, but since it's built in Rust, has anyone tried compiling it to WASM to run directly inside a browser extension?
oxirs does embeddings and GraphRAG, and full text search with Tantivy; oxirs-vec, oxirs-graphrag<p>There's an oxirs-wasm with RDF and SPARQL bindings with a query budget. Tantivy-wasm says that the release WASM bundle is 1.5 MB.<p>cool-japan/oxirs:
<a href="https://github.com/cool-japan/oxirs" rel="nofollow">https://github.com/cool-japan/oxirs</a><p>oxirs-wasm:
<a href="https://crates.io/crates/oxirs-wasm" rel="nofollow">https://crates.io/crates/oxirs-wasm</a><p>tantivy-wasm: <a href="https://github.com/phiresky/tantivy-wasm" rel="nofollow">https://github.com/phiresky/tantivy-wasm</a><p>Is there an advantage to adding an MCP local memory interface over agent instructions on how to use a rust CLI?<p>And then write Markdown documents with Google OKF-like frontmatter YAML metadata for agents that work with tokens not linked data graphs;
<a href="https://github.com/GoogleCloudPlatform/knowledge-catalog/blob/main/okf/SPEC.md" rel="nofollow">https://github.com/GoogleCloudPlatform/knowledge-catalog/blo...</a>
Can WASM use AVX512-VNNI?
Also interested.
Why not just use Qdrant? They've been integrating TurboQuant for months, works well.
What's a good embedding model and search to run locally? something fast and lightweight.
If anyone is looking to retrofit to an existing pipeline, I use similar ideas to compress vectors for job search, getting roughly 8x compression with about a 3.5% drop in quality. My experiment: <a href="https://corvi.careers/blog/vector-search-embedding-compression/" rel="nofollow">https://corvi.careers/blog/vector-search-embedding-compressi...</a>
I am not convinced that Turbovec yields better retrieval than the same amount of bits of a Matryoshka embedding.
Well. That is insane. O_O Fantastic job!
Who is this co-author called t <t@t>?
lancedb and duckdb integrations would be great...
Bloviating nonsense, 3rd time I’ve seen something like this in HN since TurboQuant came out. You don’t need float32, never did. Source: I’ve been writing on device embedding code for 4 years.
what could i use this for as part of my agentic workflow? codebase indexing? docs ?
Another vibe coded slop where they can't even spend time on Readme or documentation around code...
[dead]
[flagged]