How close can I bring vector search to where I actually need it?
I started building Itzma because I wanted to use it in my own projects. If an application needs to find the nearest vectors, can I make that a library call and keep the data in object storage? I wanted to see how far I could take that design and what I’d run into once I started using it myself.
Itzma is an open-source, embedded vector database written in Rust, with Python bindings. It’s still in beta.
The application runs the search
The application opens a namespace and queries it in-process. With an object storage backend, the bucket holds the durable data. There’s no separate vector database server to deploy.
That split raises a question I find more interesting than the initial setup. If two application processes write to the same namespace, what keeps them from overwriting each other’s changes? Itzma uses conditional writes from the object store to coordinate updates. The storage backend is part of how the database works, right down to what happens when writers overlap.
Try it from Python
Install Itzma from PyPI with pip install itzma. You can start in memory. This example inserts two vectors and asks for the closest one using cosine distance.
import itzma store = itzma.Store("memory://") schema = itzma.Schema(dim=4, metric=itzma.Metric.COSINE) docs = itzma.Namespace.create(store, "docs", schema) docs.upsert([ {"id": "a", "vector": [1, 0, 0, 0]}, {"id": "b", "vector": [0, 1, 0, 0]}, ]) hits = docs.query([0.9, 0.1, 0, 0], top_k=1) print(hits.ids) # Expected nearest ID is "a" To try the same flow against S3, point itzma.Store at s3://your-bucket/your-prefix with credentials configured. The namespace and query use the same API. For a first run with persistence on your machine, you can also point the store at a local directory.
There’s more to play with once you have a query running. Itzma supports dense and sparse retrieval, HNSW and SQ8 indexing, and filters on typed attributes.
What would you try it on?
If you have an application where you’d like to run vector search in-process, I’d like to hear what you’d need from Itzma to try it there.
Take a look at itzma.dev and start playing with it. If an example is hard to follow, a query gives you an unexpected result, or there’s a feature you’d need, tell me here or open an issue on GitHub. And if there’s something you’d like to build or fix yourself, open an issue and let’s talk through it.
GitHub and quickstart · Python package on PyPI · Rust crate on crates.io
