A vector database with semantic SQL-like filtering
Hi HN! It’s always bothered me that there’s no real equivalent of SQL WHERE for vector content. Filtering is one of the cornerstones of a modern database — but vector DBs only support either top-k sort, which is only useful for fuzzy search, or metadata filtering, which isn’t semantic. I’ve found myself wanting all the results matching my semantic query, not just k! Aside from data analysis, it's relevant if you’re trying to do any LLM reasoning: you don’t make good decisions or reach good conclusions by considering a small subset of information. So, we’ve designed a filtering primitive on…
What it does
In the maker’s words, at launch
Hi HN! It’s always bothered me that there’s no real equivalent of SQL WHERE for vector content. Filtering is one of the cornerstones of a modern database — but vector DBs only support either top-k sort, which is only useful for fuzzy search, or metadata filtering, which isn’t semantic. I’ve found myself wanting all the results matching my semantic query, not just k! Aside from data analysis, it's relevant if you’re trying to do any LLM reasoning: you don’t make good decisions or reach good conclusions by considering a small subset of information. So, we’ve designed a filtering primitive on top of vectors and assembled a demo on customer reviews from Trustpilot, Yelp, App Store, etc. You can select any brand/restaurant/app, and slice the review data however you want. The filter should find all matching documents, not just the top-k. Check it out at https://reviews.emberml.com! Not super optimized yet, and really just an exploration, but hopefully gets the point across. FAQ: - Can I try it on my own data? Sure, shoot me a message at hello [at] emberml [dot] com. - How does it work? We’ve built a custom vector-based index, and we learn a high-quality decision boundary between relevant and irrelevant vectors at query time. You can think of it as forming a few-shot classifier each time. - What’s the catch? It’s far slower and less scalable than KNN/ANN right now. But I’d rather solve quality before trying to scale up quantity; tbh I’m not satisfied with vector DB performance even at @ N=1,000. A hot take, maybe? - Why don’t you just classify the data beforehand? Unstructured data has too many degrees of freedom, so it’s hard to anticipate every search/filter a priori. Our approach is somewhat analogous to schema-on-read.
Does the same job
all alternatives →
- PAPreql – a database query language that compiles to SQL2021 · github.com · ▲38
- LSLangchain's new member-A SQL+Vector database built on ClickHouse2023 · myscale.com · ▲5
MyScale is designed for the storage and analysis of massive vector data with structured metadata. If you are eager to find a high-performance vector search using SQL queries, MyScale could be your preferred option. Thanks to the advantages of native structural database support, it provides you with a flexible filter with a WHERE clause, even JOIN when you want to jointly search vectors with filters on relevant metadata from other tables. MyScale is now open for registration and offers millions of vectors‘ free tier plan for you! (https://myscale.com/) Now you can also use…
- RCReadsql – convert SQL to most human readable format2020 · github.com · ▲43
More ai this month
the category →
I trained a 125M-parameter transformer to autocomplete piano performances in real time (~108 notes/sec on an iPhone 15). The idea is basically GitHub Copilot or Tabnine, except instead of prompting it with code, you prompt it by playing a few notes on a MIDI piano. The model then continues what you played, entirely on-device. The app is free if anyone wants to try it. Happy to answer questions about the model, training, Core ML, or the many things that didn't work.
AI · 16d ago · simedw.com
Astute▲585Automate your B2B brand going viral, with new media creators
AI · 18d ago · company-app.joinastute.com
Hey HN, Henry from Cactus here! We previously released Cactus Needle, a 14MB agentic LLM for tool call, device use, and structured extraction for phones, wearables, smart homes, small robots and microcontrollers. We got really great feedback here, and have now incorporated the suggestions to release Needle 2. The whole model is a single 14MB binary that runs a full session in 28MB of RAM; 45m parameters at 2bit compression. Needle hits 500 tokens/sec decode speed on a Raspberry Pi 5, sits between 400-1,500 tokens/sec on VR devices like Meta Quest 3S and Apple Vision Pro, and ranges…
AI · 26d ago · cactuscompute.com
