Vector

An objectstore-native vector database for approximate nearest neighbor (ANN) search. Built on SlateDB with a SPANN-style architecture: cluster centroids in memory for fast navigation, posting lists and vector data on disk for billion-scale capacity.

Features

• SPANN index: Memory-efficient ANN search with configurable centroid ratio
• Metadata filtering: Inverted indexes for hybrid queries combining similarity and predicates
• Incremental updates: LIRE-style rebalancing maintains index quality without global rebuilds
• Roaring bitmaps: Compressed posting lists for fast set operations

Quickstart

See quickstart/

Design

See rfcs/0001-storage.md for the storage design.

Docker

Build the image from the workspace root:

docker build -t opendata-vector -f vector/Dockerfile .

The container defaults to running the read-write HTTP server with config mounted at /config/vector.yaml:

docker run --rm \
  -p 8080:8080 \
  -v "$(pwd)/vector.yaml:/config/vector.yaml:ro" \
  opendata-vector

To run the read-only server instead, override the default command:

docker run --rm \
  -p 8080:8080 \
  -v "$(pwd)/reader.yaml:/config/reader.yaml:ro" \
  opendata-vector reader --config /config/reader.yaml