milvus-io/milvus-lite

Milvus Lite is a lightweight, local vector database that provides the Milvus API through a Python package for development and testing.

The tool solves the problem of needing a full Milvus deployment for prototyping and local work. It rebuilds Milvus from scratch in pure Python, replacing the C++ core with a local LSM-style storage engine. This engine includes write-ahead logging, in-memory memtables, immutable Parquet segments, and segment-level FAISS indexes. The tool supports dense vector search, sparse inner-product search, BM25 full-text search, and hybrid search, along with scalar filtering, dynamic fields, JSON fields, array fields, partitions, and aliases. It exposes both a local file-based interface through a `.db` file and an optional embedded gRPC server for multi-client development.

Milvus Lite suits developers building prototypes, working in notebooks, writing tests, or running small-scale AI applications who want to avoid the overhead of a full Milvus deployment. Code written against Milvus Lite can migrate to Milvus Standalone, Milvus Distributed, or Zilliz Cloud with minimal changes. The pure Python implementation means inspectable code and Python stack traces, which aids debugging during development. The tool is not intended for large-scale production deployments; those should use Milvus Standalone, Milvus Distributed, or Zilliz Cloud. The original C++-based milvus-lite package is not compatible with this pure-Python version, and `.db` files from the older implementation cannot be migrated automatically.

The project maintains active continuous integration across Linux on Python versions 3.10 through 3.13, macOS on Python 3.12, and Windows on Python 3.10. The codebase is written entirely in Python with core dependencies on pyarrow, numpy, faiss-cpu, and grpcio, all installed by default. The tool intentionally does not depend on pymilvus; instead, pymilvus installs Milvus Lite as its local backend.