An object-storage-native search database: vectors, full-text, and attribute filters over documents whose only durable home is an object store (a local directory or S3). A turbopuffer-inspired open-source clone, built in staged, documented commits.

AI disclaimer: this is an AI-assisted project (built with Claude and Codex). It exists to learn from, not to run your production on.

What works:

• Durable writes: a WAL with a CAS-advanced commit cursor in object storage, strongly consistent reads through the unindexed overlay, idempotency keys, conditional writes, patch/delete-by-filter, and write backpressure.
• Indexes: LSM document SSTs, delta-tiered attribute postings, BM25 full-text with rank-safe block MAXSCORE, IVF vectors with faithful RaBitQ quantization (SIMD kernels) and SPFresh-style local split/merge maintenance.
• One binary: a CLI, API, object-store-discovered queue broker, looped indexing/maintenance roles, all-in-one dev serving, operator GC dry-runs, and Prometheus /metrics endpoints for the API and queue broker.
• Backends: local filesystem for dev, plus S3-compatible stores with server-enforced conditional writes (verified against MinIO).
• Operations: namespace branch, cross-store copy, deterministic export, leased replica pinning, cache warming, and an ANN recall endpoint.

cargo run --release -- demo ./data        # tiny end-to-end demo
cargo run --release -- serve ./data       # HTTP service on 127.0.0.1:8080
cargo run --release --example usage       # library API tour
cargo run --release --example latency     # benchmark harness

The wire format tracks turbopuffer's /v2 API. With serve running (see Quick start), write a document (the namespace is created on first write):

curl -s localhost:8080/v2/namespaces/books -H 'content-type: application/json' -d '{
  "kind": "append",
  "operations": [
    {"Upsert": {"id": 1, "document": {
      "id": 1,
      "attributes": {"title": "Dune", "genre": "scifi", "rating": 4.5},
      "vectors": {"embedding": [0.9, 0.1]}
    }}}
  ]
}'

Query it. Here an approximate vector search narrowed by an attribute filter:

curl -s localhost:8080/v2/namespaces/books/query -H 'content-type: application/json' -d '{
  "kind": "single",
  "query": {
    "filter": {"Eq": {"column": "genre", "value": "scifi"}},
    "approx_vector": {"column": "embedding", "vector": [1.0, 0.0], "k": 5}
  }
}'

Hybrid search runs several rankings over one consistent snapshot with a multi-query, then fuses them client-side (for example with Reciprocal Rank Fusion). Here a vector ranking and a BM25 ranking together:

curl -s localhost:8080/v2/namespaces/books/query -H 'content-type: application/json' -d '{
  "kind": "multi",
  "query": {"queries": [
    {"approx_vector": {"column": "embedding", "vector": [1.0, 0.0], "k": 10}},
    {"text": {"column": "title", "query": "dune", "k": 10}}
  ]}
}'

The response holds one ranked result per subquery. The hybrid example (cargo run --example hybrid) shows the RRF fusion in full.

The full write/query cookbook, limits, multi-pod deployment, and the Rust library API live in the user guide.

User guide · Architecture · Build log & decisions · Contributing · Benchmarks

MIT, see LICENSE.