Tech Stack
Python
FastAPI
Docker
PostgreSQL
Description
VectorLite is a self-hostable semantic search engine built to understand approximate-nearest-neighbor search from first principles. It implements HNSW indexing in Python with NumPy-vectorized distance kernels.
Queries run a two-stage pipeline: ANN candidate retrieval followed by a hybrid rerank that blends cosine similarity with BM25 keyword scores, which noticeably improves results on short queries.
- Implemented HNSW from scratch reaching 95% recall@10 at ~2ms query latency on 1M vectors.
- Hybrid rerank (dense + BM25) improved short-query relevance ~25% over pure vector search.
- Single-container deployment with disk-backed persistence and incremental index updates.
- Published a benchmark suite comparing against FAISS and pgvector across three datasets.
Page Info
Query API & Benchmarks
REST API for upsert/query with recall-vs-latency benchmark suite against FAISS and pgvector baselines.