Indexing with vectors / vector search

LLMs have enabled a new type of search: vector search. Instead of finding a word or string, vector search turns entries and queries into vectors (arrays of numbers).

RAG search would be very useful for Atomic Assistant.

Todo

• Internal vector store
• Search UI uses vector search if available
• External embedding support (call external API)
• Internal LLM support
• E2E tests for vector search

UseCases

• Finding things by their semantic meaning (improved search)
• Searching through images, videos, audio, etc. using embeddings
• Finding relevant resources
• Injecting context in an LLM #951

Approaches

Using sled / KV store

I don't think it's possible to do meaningful nearest-neighbor search using sled's KV / BTreeMap data structure. Searching in a KV store is done with range queries over lexicographically sorted keys, and I can't see how we can turn a high-dimensional vector in a meaningful key, where close neighbors are also lexicographically close.

Using an external vector DB

The whole point of AtomicServer is that you don't need anything else. I don't want external dependencies, I want it to be a small single binary that you can just run and it gives you all you need.

OasysDB (2026: no longer mainteined)

An embeddable vector database in rust. That's the spirit!

LanceDB

• Embeddable vector search DB, seems to be really fast, also has full-text search (using tantivy). It's not OLTP
• Polle tried it, it was very slow (why?)
• Not WASM yet, maybe at some point

PolarisDB

• Zero stars
• Polle tried it, seems really fast
• Unclear if WASM