Zvec

Zvec is an open-source, in-process vector database developed by Alibaba Group and released under the Apache 2.0 license. It is designed to embed directly into applications — running as a library rather than a separate server — and the project states it has been battle-tested within Alibaba Group for production-grade, low-latency similarity search. The repository reached v0.5.0 in June 2026 and has accumulated over 11,800 GitHub stars.

What It Is

Zvec is a lightweight vector database that runs in-process alongside application code, eliminating the need for external services, network round-trips, or complex configuration. It supports dense and sparse vector embeddings, full-text search, hybrid retrieval, and structured scalar filters — all queryable in a single call. The core engine is written in C++ and exposed through official SDKs for Python, Node.js, Go, Rust, and Dart/Flutter. Its primary use cases include RAG (Retrieval-Augmented Generation) pipelines, image search, and code search.

In-Process Architecture

Unlike client-server vector databases, Zvec runs directly inside the host process. This means:

• No external daemon or container to manage
• No network latency between the application and the index
• Works in notebooks, CLI tools, servers, and edge devices
• Multiple processes can read the same collection simultaneously; writes are single-process exclusive
• Write-ahead logging (WAL) guarantees persistence even on process crash or power failure

The DiskANN index type (added in v0.5.0) keeps the bulk of the index on disk, reducing memory pressure for large-scale datasets.

Performance Benchmarks

The project publishes benchmark results on the Cohere 10M vector dataset. According to the Zvec documentation, the database indexes 10 million vectors in approximately one hour and sustains over 8,500 queries per second (QPS). The homepage describes this as "millisecond search at billion-vector scale," though the published benchmark covers the 10M dataset specifically.

SDK and Platform Coverage

Zvec ships official SDKs across five languages:

• Python (pip install zvec, Python 3.10–3.14)
• Node.js (npm install @zvec/zvec)
• Go (via the zvec-go repository)
• Rust (via the zvec-rust repository)
• Dart/Flutter (flutter pub add zvec)

Supported platforms include Linux (x86_64, ARM64), macOS (ARM64), Windows (x86_64), and RISC-V (added in v0.5.0). A companion visual tool, Zvec Studio, allows browsing data and debugging queries without writing code.

Update: v0.5.0 (June 2026)

The v0.5.0 release, published June 12, 2026, introduced several significant capabilities:

• Full-Text Search (FTS): Native keyword-based full-text search attached to any string field, with no external search engine required
• Hybrid Retrieval: A single MultiQuery can combine dense vectors, sparse vectors, scalar filters, and full-text search
• DiskANN Index: On-disk index type that drastically reduces memory usage for large datasets
• New SDKs: Official Go and Rust bindings, plus RISC-V platform support
• Zvec Studio: A visual GUI tool for data browsing and query debugging

The project roadmap is tracked publicly on GitHub issues, and the repository shows active CI/CD pipelines and ongoing community contributions.