LynseDB 0.4.0

A lightweight, server-optional, multi-end compatible, Rust-powered vector database deployable locally or remotely.

LynseDB is a high-performance, low-cost vector database that can run either embedded inside your Python process or as a standalone HTTP service. It keeps a Python-first developer experience while moving storage and search execution into a Rust backend.

Use LynseDB when you want vector search that starts as a local library, can be deployed as a service when your app grows, and does not require a heavy database cluster for common semantic search, RAG, agent memory, and multimodal retrieval workloads.

Why LynseDB

• Embedded and service-ready: the same lynse.VectorDBClient works with a local data path or a remote HTTP endpoint.
• High-performance core: dense vector search, metadata filtering, index building, and storage paths are backed by Rust.
• Low operating cost: run in-process for single-service apps and jobs, then switch to the HTTP server only when multiple workers or services need shared access.
• Retrieval beyond dense vectors: supports metadata filters, named vector fields, sparse vectors, BM25 search, and hybrid search.
• Deployment basics included: API key auth, health checks, readiness checks, Prometheus metrics, OpenAPI schema, snapshots, restore, export/import, Docker, systemd, and Kubernetes examples.

Best Fit

LynseDB is designed for teams that need a practical vector database with a small footprint:

• local semantic search inside Python apps, scripts, notebooks, and tests;
• RAG and agent-memory storage where one process can own the data directory;
• web APIs and background workers that need a shared vector service;
• multimodal records with multiple embeddings per item;
• deployments where predictable cost and simple operations matter more than running a distributed vector database cluster.

For concurrent production access from independent processes, use the HTTP server rather than sharing one embedded data directory.

Install

Python 3.9 or newer is required.

Native Linux and macOS environments are supported. Native Windows environments are not supported; on Windows, run LynseDB inside WSL 2 (Windows Subsystem for Linux) or use Docker.

pip install LynseDB

Quickstart

import numpy as np
import lynse

client = lynse.VectorDBClient(uri="./lynsedb-data")
db = client.create_database("demo", drop_if_exists=True)
collection = db.require_collection("documents", dim=4, drop_if_exists=True)

collection.add(
    ids=["intro", "guide", "note-fr"],
    vectors=[
        [0.10, 0.20, 0.30, 0.40],
        [0.11, 0.19, 0.29, 0.39],
        [0.80, 0.10, 0.20, 0.10],
    ],
    fields=[
        {"title": "LynseDB intro", "lang": "en"},
        {"title": "Vector guide", "lang": "en"},
        {"title": "French note", "lang": "fr"},
    ],
)

collection.build_index("FLAT-L2")

result = collection.search(
    np.array([0.10, 0.20, 0.30, 0.40], dtype=np.float32),
    k=2,
    where="lang = 'en'",
    return_fields=True,
)

print(result.to_list())

One API, Two Deployment Modes

Embedded mode

Use embedded mode when one Python process owns the data directory. It avoids a network hop and is the lowest-cost way to add vector search to a local app, notebook, test suite, job, or small service.

client = lynse.VectorDBClient(uri="./data")

Service mode

Use service mode when multiple processes, web workers, or applications need to share the same database.

lynse serve --host 0.0.0.0 --port 7637 --data-dir ./server-data

client = lynse.VectorDBClient("http://127.0.0.1:7637")

With API key authentication:

lynse serve --host 0.0.0.0 --port 7637 --data-dir ./server-data --api-key your_key

client = lynse.VectorDBClient("http://127.0.0.1:7637", api_key="your_key")

Health, readiness, metrics, and OpenAPI endpoints are available in service mode:

curl http://127.0.0.1:7637/healthz
curl http://127.0.0.1:7637/readyz
curl http://127.0.0.1:7637/metrics
curl http://127.0.0.1:7637/openapi.json

Retrieval Features

• Dense vector search with flat, HNSW, IVF, DiskANN, and quantized index families.
• Standard SQL-style metadata filtering through where expressions.
• Named vector fields for multimodal records, such as text and image embeddings on the same item.
• Sparse vector search for feature-weight retrieval.
• BM25 search over metadata fields.
• Hybrid search with vector and text candidates.
• ResultView return objects with NumPy arrays plus list, JSON, and dataframe conversion helpers.

Indexing

Start with a flat index as a correctness baseline:

collection.build_index("FLAT-L2")

Move to HNSW or IVF when latency matters, DiskANN when memory pressure matters, and quantized variants when you want a smaller memory or disk footprint:

collection.build_index("HNSW-L2")
collection.build_index("IVF-L2", n_clusters=256)
collection.build_index("DiskANN-L2")
collection.build_index("FLAT-IP-SQ8")
collection.build_index("FLAT-L2-PQ")

See the indexing guide for metric names, nprobe tuning, binary indexes, and quantized index variants.

Docker

docker run -p 7637:7637 -v lynsedb-data:/data birchkwok/lynsedb:latest
docker run -p 7637:7637 -e LYNSE_API_KEY=your_key -v lynsedb-data:/data birchkwok/lynsedb:latest

On Windows, use this Docker image or install/run LynseDB from a Linux environment in WSL 2.

Deployment examples are included in:

• Docker Compose
• systemd
• Kubernetes

Documentation

• Learning path
• Core concepts
• Quickstart
• Connect and deploy
• Databases and collections
• Add vectors
• Search and filter
• Metadata filter cookbook
• Indexing guide
• Named, sparse, and hybrid search
• Build a RAG workflow
• Performance tuning
• Backup and maintenance
• Troubleshooting
• Client API
• ResultView

Stability Notes

LynseDB is still evolving. Pin package and server image versions for deployments, and test migrations before upgrading. For concurrent production access, prefer the HTTP server over sharing one local data directory across independent Python processes.