Vector Database
An HNSW-based in-memory vector database in Rust with configurable distance metrics, quantization, segmented indexing, and concurrent search.
Problem
Similarity search systems need low-latency nearest-neighbor retrieval while balancing memory efficiency, persistence, and concurrent access.
Approach
I implemented an HNSW-powered vector engine with L2 and cosine metrics, scalar quantization, and memory-mapped fixed-capacity storage. The index model combines immutable and mutable segments with flush support, while search is parallelized with Rayon.
Outcome
- Delivered thread-safe concurrent search with persistent JSON-backed state.
- Added segmented index architecture to support mutable ingestion and stable query behavior.
- Built a configurable vector search foundation suitable for retrieval-heavy workloads.