VectorLite

A tiny, in-process Rust vector store with built-in embeddings for sub-millisecond semantic search.

VectorLite is a high-performance, in-memory vector database optimized for AI agent and edge workloads.
It co-locates model inference (via Candle) with a low-latency vector index, making it ideal for session-scoped, single-instance, or privacy-sensitive environments.

Why VectorLite?

VectorLite trades distributed scalability for deterministic performance, perfect for use cases where latency mattters more than millions of vectors.

When to Use It

Quick Start

Run from Source

cargo run --bin vectorlite -- --port 3001

# Start with preloaded collection
cargo run --bin vectorlite -- --filepath ./my_collection.vlc --port 3001

Run with Docker

With default settings:

docker build -t vectorlite .
docker run -p 3001:3001 vectorlite

With a different embeddings model and memory-optimized HNSW:

docker build \
  --build-arg MODEL_NAME="sentence-transformers/paraphrase-MiniLM-L3-v2" \
  --build-arg FEATURES="memory-optimized" \
  -t vectorlite-small .

HTTP API Overview

Index Types

See Hierarchical Navigable Small World.

Configuration profiles for HNSW

cargo build --features memory-optimized

Similarity Metrics

• Cosine: Default for normalized embeddings, scale-invariant
• Euclidean: Geometric distance, sensitive to vector magnitude
• Manhattan: L1 norm, robust to outliers
• Dot Product: Raw similarity, requires consistent vector scaling

Rust SDK Example

use vectorlite::{VectorLiteClient, EmbeddingGenerator, IndexType, SimilarityMetric};
use serde_json::json;

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let client = VectorLiteClient::new(Box::new(EmbeddingGenerator::new()?));

    client.create_collection("quotes", IndexType::HNSW)?;
    
    let id = client.add_text_to_collection(
        "quotes", 
        "I just want to lie on the beach and eat hot dogs",
        Some(json!({
            "author": "Kevin Malone",
            "tags": ["the-office", "s3:e23"],
            "year": 2005,
        }))
    )?;

    let results = client.search_text_in_collection(
        "quotes",
        "beach games",
        3,
        SimilarityMetric::Cosine,
    )?;

    for result in &results {
        println!("ID: {}, Score: {:.4}", result.id, result.score);
    }

    Ok(())
}

Testing

Run tests with mock embeddings (CI-friendly, no model files required):

cargo test --features mock-embeddings

Run tests with local models:

cargo test

Download ML Model

This downloads the BERT-based embedding model files needed for real embedding generation:

huggingface-cli download sentence-transformers/all-MiniLM-L6-v2 --local-dir models/all-MiniLM-L6-v2

The model files must be present in the ./models/{model-name}/ directory with the required files:

• config.json
• pytorch_model.bin
• tokenizer.json

Using a different model

You can override the default embedding model at compile time using the custom-model feature:

DEFAULT_EMBEDDING_MODEL="sentence-transformers/paraphrase-MiniLM-L3-v2" cargo build --features custom-model

DEFAULT_EMBEDDING_MODEL="sentence-transformers/paraphrase-MiniLM-L3-v2" cargo run --features custom-model

License

Apache 2.0 License - see LICENSE for details.

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.