Lib.rs

Victor

Web-optimized vector database (written in Rust).

Features

1. Rust API (using native filesystem, or a transient in-memory filesystem)
2. Web API (Using the Private Origin File System)
3. Very efficient vector storage format
   1. For a vector with 1536 dimensions, our representation consumes 1.5 KB, while naively encoding with JSON would consume 20.6 KB.
4. PCA for vector compression when storage space is low

JS Example

Installation

npm install victor-db

Usage

import { Db } from "victor";

const db = await Db.new();

const content = "My content!";
const tags = ["these", "are", "tags"];
const embedding = new Float64Array(/* your embedding here */);

// write to victor
await db.insert(content, embedding, tags);

// read the 10 closest results from victor that are tagged with "tags"
// (only 1 will be returned because we only inserted one embedding)
const result = await db.search(embedding, ["tags"], 10);
assert(result[0].content == content);

// clear database
await db.clear();

See www/ for a more complete example, including fetching embeddings from OpenAI.

Rust Example

Installation

cargo add victor-db

Usage

The Rust API can automatically create embeddings for you with fastembed-rs's default model (currently BAAI/bge-small-en-v1.5).

use std::path::PathBuf;

use victor_db::native::Db;

let _ = std::fs::create_dir("./victor_test_data");
let mut victor = Db::new(PathBuf::from("./victor_test_data"));

victor.clear_db().await.unwrap();

victor
    .add(
        vec!["Pineapple", "Rocks"], // documents
        vec!["Pizza Toppings"],     // tags (only used for filtering)
    )
    .await;

victor
    .add_single("Cheese pizza", vec!["Pizza Flavors"])
    .await; // Add another entry with no tags

// read the 10 closest results from victor that are tagged with "Pizza Toppings"
// (only 2 will be returned because we only inserted two embeddings)
let nearest = victor
    .search("Hawaiian pizza", vec!["Pizza Toppings"], 10)
    .await
    .first()
    .unwrap()
    .content
    .clone();
assert_eq!(nearest, "Pineapple".to_string());

This example is also in the /examples directory. If you've cloned this repository, you can run it with cargo run --example native_filesystem.

Hacking

1. Victor is written in Rust, and compiled to wasm with wasm-pack.

   Install wasm pack with cargo install wasm-pack or npm i -g wasm-pack (https://rustwasm.github.io/wasm-pack/installer/)

2. Build Victor with wasm-pack build --target web

3. Set up the example project, which is in www/.

   If you use nvm, you can just run cd www/ && nvm use

   Then, npm i.

4. From www/, start the example project with npm run start.

Architecture

Relevant code at src/packed_vector.rs.

Us

Sam Hall

Andre Popovitch