Home
S3-native vector search engine.
Object storage is the source of truth. Nodes are stateless.
- S3-Native — All data lives in object storage (S3, GCS, Azure Blob, local filesystem). No local databases. Nodes are disposable.
- IVF Indexing — IVF-Flat, IVF-SQ8 (4x compression), IVF-PQ (16-32x compression), and Hierarchical ANN indexes.
-
BM25 Full-Text Search — Inverted indexes with configurable tokenization, stemming, and multi-field
rank_byexpressions. Combine vector and keyword search. - Bitmap Pre-Filters — RoaringBitmap indexes for sub-millisecond attribute filtering before vector search.
- Strong Consistency — WAL-based writes with CAS (compare-and-swap) on manifest. Strong or eventual consistency per query.
- Formally Verified — 15 TLA+ specifications verify correctness of compaction, multi-writer leases, BM25 consistency, and more.
- 198 unit tests across all modules
- 15 TLA+ specifications with 26,000+ explored states
- 6 stress tests and 6 end-to-end performance tests
- Written in Rust (1.84+), async with Tokio, HTTP via Axum 0.7
- Apache 2.0 license
| Section | Description |
|---|---|
| Getting Started | Docker quickstart, first API calls |
| API Reference | Complete HTTP API with JSON schemas |
| Configuration | All config options, env vars, TOML format |
| Filters | Filter query language with examples |
| Full-Text Search | BM25 ranking, tokenization, rank_by expressions |
| Python SDK | Python client guide |
| TypeScript SDK | TypeScript client guide |
| Architecture | System design, data flow, S3 key structure |
| Indexing | Vector indexing internals (IVF, SQ8, PQ, H-ANN) |
| WAL & Compaction | Write-ahead log, manifest, compaction pipeline |
| Deployment | Docker, EC2, production tuning |
| Contributing | Build, test, CI pipeline, code style |
| Formal Verification | TLA+ specs and findings |
| Metrics Reference | All Prometheus metrics |