Zvec is an open-source, lightweight, in-process vector database from Alibaba built on top of Proxima, Alibaba's production vector search engine. It embeds directly into applications as a library rather than running as a separate service, enabling millisecond-level similarity search across billions of vectors with zero operational overhead. ## Why an In-Process Vector Database The vector database landscape is crowded with server-based solutions like Milvus, Qdrant, Pinecone, and Weaviate. These systems work well for large-scale deployments but introduce operational complexity: servers to manage, network latency to contend with, and infrastructure costs that scale with query volume. Zvec takes a fundamentally different approach. By embedding the vector search engine directly into the application process, it eliminates the network hop entirely. There are no servers to configure, no ports to manage, and no client-server protocol overhead. This makes it ideal for edge devices, notebooks, CLI tools, and applications where simplicity and low latency are paramount. ## Core Architecture Zvec is built in C++ (81.5% of the codebase) for maximum performance, with Python and Node.js bindings provided through SWIG. The underlying search algorithms come from Proxima, which has been battle-tested in Alibaba's production systems handling massive-scale similarity search workloads. The library supports both dense and sparse vector embeddings with multi-vector query capabilities. This dual-vector support means applications can combine traditional embedding-based semantic search with sparse representations like BM25 or SPLADE in a single query, enabling hybrid retrieval strategies. ## Hybrid Search Capabilities Beyond vector similarity, Zvec supports structured filtering that can be combined with vector queries. This means applications can run queries like "find the most semantically similar documents that were also published in the last 30 days and belong to category X." The filtering happens at the search level rather than as a post-processing step, maintaining query efficiency. ## Installation and Platform Support Zvec is available through pip for Python (3.10-3.12) and npm for Node.js. The library supports Linux (x86_64 and ARM64) and macOS (ARM64), covering the most common development and deployment platforms. Installation is a single command: `pip install zvec` or `npm install @zvec/zvec`. From there, creating a collection, inserting vectors, and running similarity searches takes just a few lines of code. ## Performance Characteristics The project claims millisecond-level search times across billions of vectors, leveraging Proxima's optimized indexing structures. Being in-process eliminates network serialization overhead, which can be significant for high-throughput applications making thousands of queries per second. The C++ core ensures minimal memory overhead per vector, and the library manages its own memory allocation without relying on garbage collection in the host language. ## Use Cases Zvec's in-process design makes it particularly suited for several scenarios. Jupyter notebooks and data science workflows benefit from instant setup without external dependencies. Edge devices and embedded systems can run similarity search without network connectivity. CLI tools can include vector search capabilities without requiring users to run a separate service. And rapid prototyping becomes easier when the entire vector search stack is a single pip install away. ## Comparison with Alternatives Compared to server-based solutions like Milvus or Qdrant, Zvec trades distributed scalability for simplicity and latency. Compared to FAISS, which is also an in-process library, Zvec provides a higher-level API with built-in persistence and filtering capabilities. The Proxima engine underlying Zvec has been production-tested at Alibaba's scale, giving it credibility that newer projects lack. ## Limitations As an in-process library, Zvec is limited to the memory and CPU of the host machine. It does not support distributed queries across multiple nodes, making it unsuitable for truly massive-scale deployments that exceed single-machine capacity. The platform support is currently limited to Linux and macOS, excluding Windows. Python support covers versions 3.10 through 3.12, which may require version management for some users. Documentation, while functional, is still maturing as the project is in its early releases (v0.2.0).