@zvec/zvec

Implements approximate nearest neighbor (ANN) search using in-process indexing structures that avoid network round-trips and external database dependencies. The engine builds spatial index structures (likely HNSW or similar graph-based ANN algorithms) over vector embeddings stored in memory, enabling sub-millisecond similarity queries without serialization overhead. Queries return ranked results by cosine/L2 distance without requiring cloud connectivity or managed service infrastructure.

Supports attaching arbitrary metadata (tags, categories, timestamps, source URLs) to vectors and filtering results by metadata predicates before or after similarity ranking. Enables hybrid search patterns combining vector similarity with structured filtering (e.g., 'find similar documents from the last 30 days in category X'). Metadata is stored alongside vectors in the index structure, allowing efficient pre-filtering to reduce search space.

Supports adding vectors to the index in batches or individually without rebuilding the entire index structure. Uses incremental insertion algorithms (likely HNSW layer insertion or similar) that maintain index quality while adding new vectors. Batch operations are optimized to amortize insertion overhead across multiple vectors, reducing per-vector insertion cost compared to individual inserts.

Supports multiple distance metrics (cosine similarity, Euclidean L2, dot product, Hamming distance) for computing vector similarity, allowing users to choose the metric that best matches their embedding model and use case. Metrics are pluggable at index creation time and applied consistently across all queries. Similarity scores are normalized and returned alongside results for ranking and threshold-based filtering.

Stores vectors in a compact in-memory format with optional quantization or compression to reduce memory footprint. Uses typed arrays (Float32Array) for efficient storage and may support lower-precision formats (float16, int8) for approximate storage with reduced memory overhead. Compression trades query accuracy for memory efficiency, useful for large collections on memory-constrained environments.

Provides specialized indexing and search for code snippets and source files by understanding code structure (functions, classes, imports) and language-specific semantics. Embeds code at multiple granularities (file, function, class level) and enables searching by intent (e.g., 'find functions that validate email addresses') rather than keyword matching. Supports multiple programming languages with language-specific tokenization and embedding strategies.

Loads vector indexes from disk using memory-mapping (mmap) to avoid copying entire indexes into memory, instead mapping file pages directly to virtual memory. Enables loading indexes larger than available RAM by paging in vectors on-demand. Zero-copy access patterns minimize memory overhead and startup time, particularly beneficial for large pre-computed indexes that are loaded once and queried many times.