HNSW graphs
Vectors join a layered navigable graph on insert. Search descends from a sparse top layer to a dense base layer, following the shortest path a greedy walk can find.
Embedded vector search · index 3.0
Jasper
A vector search engine that links into your process instead of running beside it. HNSW graphs, tiered quantization, and hybrid keyword scoring, in a library with no server to operate.
Jasper ships as a linkable library, not a service. You open an index file the way you'd open a database file, insert vectors as your application produces them, and query it in the same process — no network hop, no separate deployment to keep patched and paged for.
Under the surface it is a hierarchical navigable small world graph over your vectors, with an optional inverted index sitting beside it for exact keyword terms. The manual below walks both halves: how the graph is built and searched, how vectors are compressed into cheaper tiers without losing what makes them useful, and how a query blends keyword and vector signal into one ranked list.
Read the quickstart Start with index anatomy
Press Ctrl K to jump straight to a concept.
What lives inside one index file
Quantization tiers
Store vectors at the precision each collection actually needs — full float32, scalar int8, or product-quantized codes — and rerank the shortlist at full precision.
Hybrid scoring
Fuse a keyword score against the same postings you already index with a vector score from the graph, so exact terms and semantic neighbors both count.
The concepts, in the order they compose
Six chapters, read start to end or jumped to directly — each one builds on the vocabulary the last one set down.
- Index Anatomy 2 min What a .jasper file actually contains: manifest, vector store, graph store, and write-ahead log.
- Building the HNSW Graph 2 min Layers, M, and ef_construction: how vectors join a hierarchical navigable small world graph.
- Quantization Tiers 2 min Trading precision for memory with scalar and product quantization, and reranking to get it back.
- Hybrid Keyword Scoring 2 min Fusing an inverted-index keyword score with vector similarity into one ranked list.
- Query Tuning 2 min The ef_search, oversampling, and filter knobs that trade recall against latency at query time.
- Persistence and Recovery 2 min The write-ahead log, checkpoints, and what happens the moment Jasper reopens after a crash.