Back to selected work
Vector Search Studio
Semantic search over Supabase docs with pgvector and a local embedder. No API keys.
Outcome
Semantic search over Supabase docs with no API keys.
- Year
- 2026
- Role
- Design and build
- Status
- Independent
- Stack
- Next.js, TypeScript, pgvector, HNSW, Vitest
The problem
Keyword search misses intent, and hosted embeddings add keys and cost.
What I built
A 384-dimension random-projection embedder, a vector(384) column, an HNSW cosine index, SQL distance queries, an App Router UI, and tests in CI.
Query path
- Local embedder
- vector(384)
- HNSW index
- SQL distance
Decisions
No API keys.
Ranking is covered by tests.
Tradeoff
Weaker than a trained embedding model.
What comes next
A real embedder, reranking, and recall measurement.