milvus-io/milvus-lite

Milvus Lite – Quick Overview

What it is: Milvus Lite is a pure‑Python, local‑only version of the Milvus vector database. It lets you store vectors, scalar fields, and text, and run similarity search, filtering, and hybrid (vector + text) queries without needing the full Milvus server or any C++ components.

Why it matters: Developers often need a lightweight, easy‑to‑install database for notebooks, unit tests, or small prototypes. Milvus Lite mimics the Milvus pymilvus API, so code written against it can later be switched to a production Milvus deployment (stand‑alone, distributed, or Zilliz Cloud) with minimal changes.

Key capabilities:

• Local storage – data lives in a single .db file (or a directory) using an LSM‑tree style engine with WAL, memtables, and Parquet segments.
• Vector search – dense vectors via FAISS indexes (HNSW, IVF, etc.) and sparse vectors with exact inner‑product or BM25 scoring.
• Full‑text & hybrid search – built‑in BM25 analyzer, optional Jieba for Chinese, and the ability to combine vector and text scores.
• Rich schema – support for many scalar types, dynamic fields, JSON, arrays, geometry (WKT), and functions like OpenAI text‑embedding.
• gRPC adapter – runs a local Milvus‑compatible server so multiple Python processes can share the same database.
• Python‑only – all core logic is written in Python, giving clear stack traces and easy inspection.

Typical usage:

from pymilvus import MilvusClient
client = MilvusClient("./demo.db")
client.create_collection("docs", dimension=4)
client.insert("docs", [{"id":1, "vector":[0.1,0.2,0.3,0.4], "text":"python tutorial"}])
results = client.search("docs", data=[[0.1,0.2,0.3,0.4]], limit=2)

You can also start a server (milvus-lite server) and connect via uri="http://127.0.0.1:19530" or use the embedded engine API directly (from milvus_lite import MilvusLite).

Limitations: single‑process writes, no authentication, no distributed scaling, no binary/float16 vectors, and some advanced Milvus features (e.g., product‑quantization indexes, spatial indexes) are missing.

---

Repository Summary (from the README)

Aspect	Details
License	Apache 2.0
Language	Python 3.10+
Installation	pip install -U milvus-lite (or via pymilvus[milvus-lite])
Core deps	pyarrow, numpy, faiss‑cpu, grpcio
Main components	Adapter (gRPC), Engine (storage, search, index, analyzer, function, schema)
Supported data types	Dense & sparse vectors, BOOL, integers, FLOAT, DOUBLE, VARCHAR, JSON, ARRAY, GEOMETRY, TIMESTAMPTZ
Search metrics	COSINE, L2, IP, BM25
Indexes	HNSW, IVF, FLAT, BRUTE_FORCE, sparse inverted, scalar inverted
Features	CRUD, partitions, aliases, snapshots, hybrid search, reranking, full‑text, geometry filters
Running modes	Local .db file, embedded engine API, local gRPC server
Docs & tests	Design docs in docs/, extensive CI across OSes and Python 3.10‑3.13
Typical users	Notebook developers, CI tests, small‑scale AI apps, anyone needing Milvus‑compatible API without a full server

---

Who Might Benefit?

• Data scientists building proof‑of‑concept similarity search pipelines in Jupyter.
• Engineers writing unit or integration tests for code that will eventually talk to a real Milvus cluster.
• Students learning vector databases and wanting to experiment locally without Docker or heavy installations.

---

Quick Start Checklist

1. pip install "pymilvus[milvus-lite]"
2. Create a client with a path to a .db file.
3. Define a collection schema (or let Milvus create one automatically).
4. Insert vectors / text / scalar fields.
5. Run search, query, or delete as needed.
6. For multi‑process work, start milvus-lite server --data-dir ./data --port 19530 and point clients to the URI.

---

All information above is taken directly from the project's README; no external assumptions have been added.