Quiver

A fast, embeddable vector database written from scratch in Rust.

Store vectors with JSON metadata, find their nearest neighbours with HNSW, exact or k-d tree search, and filter as you go. Run it as a server with REST and gRPC, or embed it in your Rust program.

$git clone https://github.com/HilthonTT/quiver && cd quiver && cargo run --release -p quiver-server

Get started Star on GitHub

Features

A whole vector database, no dependencies to run.

One binary serves REST on port 6900 and gRPC on 6901. Collections, indexes, the write-ahead log and the snapshot archives are all Quiver's own code.

Three indexes

HNSW for large collections, exact flat search, and a k-d tree for exact search in low dimensions.

Five metrics

Cosine, Euclidean, dot product, Manhattan and Hamming. Distances are always lower = closer.

Filtered search

MongoDB-style metadata filters ($gte, $in, $or, nested fields), applied during the index traversal.

Data operations

Merge or replace metadata, count and delete by filter, search by id, recommend, distance thresholds.

Durable storage

A checksummed write-ahead log with checkpoints by default, RocksDB behind a feature flag, or memory.

Snapshots

SHA-256-verified .tar.gz backups you can download, upload and restore, with retention and a schedule.

API keys

REST and gRPC behind read-only and read-write keys, compared in constant time.

Embeddable

Run it as a server, or link the quiver-db crate and keep everything in-process.

Python and Rust clientsTyped sync and async Python with numpy support, and an async Rust client.A terminal UIBrowse collections, insert, search and manage snapshots, with optional text embeddings.

In code

From Python, from curl, or in-process.

The same collection, three ways in. Every client speaks the same JSON filters and returns the same lower-is-closer distances.

Pythonpython

# pip install ./clients/python
from quiver import Client

client = Client("http://localhost:6900")
books = client.get_or_create_collection("books", dimension=3)
books.upsert([
    ("dune", [0.9, 0.1, 0.0], {"year": 1965}),
    ("hobbit", [0.1, 0.9, 0.2], {"year": 1937}),
])
for hit in books.search([1.0, 0.0, 0.0], k=1, filter={"year": {"$gt": 1950}}):
    print(hit.id, hit.distance)  # dune 0.006116271018981934

RESTsh

# create a collection, add a point, search
curl -X POST localhost:6900/collections \
  -H 'content-type: application/json' \
  -d '{"name": "places", "dimension": 2, "metric": "euclidean"}'

curl -X POST localhost:6900/collections/places/points \
  -H 'content-type: application/json' \
  -d '{"points": [{"id": "brussels", "vector": [4.35, 50.85]}]}'

curl -X POST localhost:6900/collections/places/search \
  -H 'content-type: application/json' \
  -d '{"vector": [5.5, 50.5], "k": 1}'

Embeddedrust

// cargo add quiver-db --git https://github.com/HilthonTT/quiver
use quiver_db::{CollectionConfig, Database, Point};
use serde_json::json;

let db = Database::open("./quiver-data")?;
let books = db.create_collection("books", CollectionConfig::new(3))?;
books.upsert([
    Point::new("dune", vec![0.9, 0.1, 0.0])
        .with_metadata(json!({"year": 1965})),
])?;
for hit in books.search(&[1.0, 0.0, 0.0], 1)? {
    println!("{} {}", hit.id, hit.distance);
}

Walk through the quick start

Clone it, run it, search.

One command builds and starts the server with a local data directory. The docs cover the rest, from filters to snapshots to API keys.

$git clone https://github.com/HilthonTT/quiver && cd quiver && cargo run --release -p quiver-server

Read the docs