Neural Search Engine
A hybrid search system combining dense vector retrieval with sparse lexical matching, designed for technical documentation at scale.
AI
Problem
Traditional keyword search fails on technical documentation. Engineers search with intent — "how to handle backpressure in async pipelines" — but lexical matching returns noise.
Approach
Built a hybrid retrieval system that fuses dense and sparse signals at query time:
- Dense path: Sentence-transformer embeddings indexed in HNSW (via Qdrant)
- Sparse path: BM25 over tokenized documents
- Fusion layer: Reciprocal Rank Fusion (RRF) with dynamic weights
Results
- Hybrid retrieval: nDCG@10 of 0.82 (vs. 0.71 dense-only)
- Query latency: p50 = 32ms, p99 = 118ms
- 40% reduction in "no result found" feedback