Zvec: Alibaba's Open-Source "SQLite for Vector Databases" — Complete Developer Guide

A deep-dive into the in-process vector database that's taking the AI world by storm. 11.5k GitHub stars and counting.

What Is Zvec?

Zvec is an open-source, in-process vector database built by Alibaba's Tongyi Lab and released in February 2026. Think of it as the SQLite of vector databases — it runs entirely inside your application process. No server. No daemon. No network calls. Just pip install zvec and you're running production-grade vector search.

It's built on Proxima, Alibaba's battle-tested internal vector search engine that handles billions of queries daily across Alibaba's search, recommendation, and advertising systems.

License: Apache-2.0
GitHub Stars: ~11.5k
Latest Release: v0.5.0 (June 12, 2026)

Why Does This Matter?

Vector databases power modern AI applications. When an app finds products similar to what you bought, or documents related to your question, a vector database is working in the background.

The problem: Good vector databases have been either:

Expensive (Pinecone starts at $70/month)
Complex to operate (Chroma and Weaviate need infrastructure and setup time)

Zvec's answer: One line of code. Zero configuration. Runs anywhere your code runs.

Installation

Python

pip install zvec
# Requires Python 3.10–3.14

Node.js

npm install @zvec/zvec

Go

go get github.com/alibaba/zvec-go

Rust

[dependencies]
zvec = "0.5"

Flutter/Dart

flutter pub add zvec

Zvec Studio (Visual Tool)

pip install zvec-studio

Supported Platforms: Linux (x86_64, ARM64) · macOS (ARM64) · Windows (x86_64) · Android · iOS · RISC-V

Quickstart: Your First Zvec App in 60 Seconds

import zvec

# 1. Define your schema
schema = zvec.CollectionSchema(
    name="movies",
    vectors=zvec.VectorSchema("embedding", zvec.DataType.VECTOR_FP32, 128),
)

# 2. Create the collection (it's just a local directory)
collection = zvec.create_and_open(path="./movies_db", schema=schema)

# 3. Insert documents
collection.insert([
    zvec.Doc(id="movie_1", vectors={"embedding": [0.1, 0.2, ..., 0.4]}),
    zvec.Doc(id="movie_2", vectors={"embedding": [0.2, 0.3, ..., 0.1]}),
])

# 4. Search by vector similarity
results = collection.query(
    zvec.VectorQuery("embedding", vector=[0.4, 0.3, ..., 0.1]),
    topk=10
)

print(results)
# Returns: list of {'id': str, 'score': float, ...} sorted by relevance

That's it. No servers, no containers, no cloud accounts.

What v0.5.0 (June 12, 2026) Brings

The latest release is a massive upgrade:

1. Full-Text Search (FTS) 🔍

Native keyword-based full-text search — attach an FTS index to any string field and query it with natural-language or structured expressions. No Elasticsearch or external engine needed.

# Create FTS index on a string field
collection.create_index("title", zvec.FTSIndexParam())

# Query with full-text
results = collection.query(
    zvec.TextQuery("title", "quantum computing"),
    topk=10
)

2. Hybrid Retrieval 🧬

Combine full-text search + vector similarity + scalar filters + sparse vectors in a single MultiQuery. This is where Zvec really shines — you get the precision of semantic search with the keyword-matching power of traditional search.

results = collection.query(
    zvec.MultiQuery([
        zvec.VectorQuery("embedding", vector=[0.1, 0.2, ...]),
        zvec.TextQuery("title", "machine learning"),
        zvec.Filter("year > 2020"),
    ]),
    topk=20,
    rerank=True  # Built-in RRF reranking
)

3. DiskANN Index 💾

A new on-disk index that keeps the bulk of the index on disk instead of RAM. This dramatically cuts memory costs for billion-scale datasets, making large-scale ANN search feasible on memory-constrained machines.

4. Zvec Studio 🎨

A visual management tool to browse data, test queries, and manage schemas without writing code. Install via pip install zvec-studio or download the desktop app.

5. New Official SDKs

Go SDK (zvec-go): cgo bindings with prebuilt libraries for all major platforms
Rust SDK (zvec-rust): Safe, idiomatic bindings with RAII resource management and builder APIs

Performance Benchmarks

Tested on the Cohere 10M vector dataset:

Metric	Value
Total Vectors Indexed	10,000,000
Index Build Time	~1 hour
Queries per Second (QPS)	8,500+
Search Latency	Milliseconds at billion-vector scale

Key Features Deep Dive

In-Process Architecture

Zvec runs directly in your application's memory space. No RPC, no network overhead, no daemon management. This means:

Zero latency from network calls
No infrastructure to maintain
Works offline — perfect for edge devices and on-device AI

Dense + Sparse Vectors

Support both dense embeddings (like OpenAI's text-embedding-3) and sparse embeddings (like SPLADE), with multi-vector queries.

Durable Storage

Write-Ahead Logging (WAL) guarantees persistence. Your data survives process crashes and power failures.

Concurrent Access

Multiple processes can read the same collection simultaneously. Writes are single-process exclusive — consistent and safe.

Filtered & Grouped Search

# Vector search with filters
results = collection.query(
    zvec.VectorQuery("embedding", vector=[...]),
    filter=zvec.Filter("price < 50 AND category = 'electronics'"),
    topk=10
)

# GROUP BY style search
results = collection.query(
    zvec.VectorQuery("embedding", vector=[...]),
    group_by="category",
    topk_per_group=5
)

Real-World Use Cases

1. Retrieval-Augmented Generation (RAG)

Enhance LLM responses by retrieving relevant context from your knowledge base — all running locally.

2. Image Search

Find visually or semantically similar images at scale. Perfect for e-commerce product discovery.

3. Code Search

Find code snippets by describing wh