Stable documents · low write rate
Internal docs, standards, product manuals
A bounded document library that grows in batches, engineering wikis, policy libraries, regulatory standards, technical reference. High read volume, low write rate.
A large, stable document library costs almost nothing to keep searchable, and you can adopt a better model any time without rebuilding it.
01
Bulk ingest on a schedule
Your pipeline chunks documents and posts them to the bulk ingest endpoint. Rebuilds happen weekly or on document update.
02
GraphANN™ holds the graph + compressed text
No embeddings persisted. Storage scales with your content, not your choice of embedding model.
03
Application queries directly
Search returns ranked passages with source metadata. Drop-in replacement for keyword search or pgvector.
Best fit when
- Document library spans thousands to millions of records
- Updates arrive in batches, not as continuous streams
- You care about storage cost or model lifecycle, not raw ingest QPS
- You want to swap embedding models without rebuilding the world
Wrong choice when
- You need transactional consistency between document writes and search
- Your document library is already small enough that a simple search works fine
- You need rich SQL-style joins on document properties at query time
Up to 95%
less storage at scale
Bulk
ingest endpoint
Single
binary deploy