The vector database that refuses to be a microservice

What it does Zvec is a C++ vector search engine that compiles into your application process, offering dense and sparse vector storage, hybrid search with structured filters, and write-ahead logging for durability. It exposes language bindings for Python, Node.js, and Dart/Flutter, aiming to keep latency low by eliminating network hops to a separate database server. Alibaba describes it as battle-tested within its own production environment.

The interesting bit Most vector databases want you to deploy a cluster or pay for a managed service; Zvec treats similarity search as a library call you make from a notebook, CLI tool, or mobile app. Its latest release even ships a Dart/Flutter SDK with FFI bindings, targeting edge and mobile use cases rather than just data-center servers.

Key highlights

• Supports dense and sparse vectors, plus multi-vector queries in a single call.
• Hybrid search mixing semantic similarity with structured metadata filters.
• Durability through write-ahead logging (WAL), with recovery from process crashes or power loss.
• Concurrent multi-process reads, though writes are restricted to a single process.
• Cross-platform C++ core with bindings for Python, Node.js, and Dart/Flutter.

Caveats

• The supported-platform list covers Linux, macOS (ARM64), and Windows (x86_64); iOS support is mentioned in the v0.4.0 release notes but not yet reflected in the main compatibility matrix.
• Version 0.4.0 patched bugs affecting SQ8 quantizer recall and sparse-vector index ordering, suggesting quantization and sparse indexing are still stabilizing.

Verdict Developers building RAG pipelines, local LLM memory, or edge-device search who want to avoid operating yet another network service should take a look. If you need a distributed, multi-writer vector store with complex replication, this is not your tool.