bicardinal/brinicle

Implements a disk-persisted approximate nearest neighbor index with support for dynamic operations (insert, upsert, delete) without rebuilding, achieving FAISS/hnswlib-comparable speed while keeping the index entirely off-heap. Uses a graph-based approach optimized for flash storage access patterns and sequential I/O to minimize memory footprint on resource-constrained deployments. Provides Python bindings and targets production systems requiring predictable latency with high recall on vector similarity queries.

Available on PyPI.