Open-source vector infrastructure

Vector search,
engineered from first principles.

Axion is a durable Rust vector engine with native HNSW, RocksDB, REST and gRPC APIs, a typed Python client, and deployment paths that are validated—not merely documented.

RocksDB default Official ANN-Benchmarks kind-validated
QUERY / 384D nearest_neighbors(k=10)
HNSW / READY

Move your pointer to perturb the query vector.

THE PARTS THAT MATTER

01 Predictable latency
02 Explicit trade-offs
03 Durable state
04 Open internals

01 / THE ENGINE

A short, inspectable path from vector to answer.

API boundaries validate. The collection engine plans and synchronizes. Exact or HNSW indexes score candidates. RocksDB owns durable truth.

  1. 01ClientsREST · gRPC · Python
  2. 02API boundaryTyped validation
  3. 03Query engineFilter · plan · merge
  4. 04Exact / HNSWDeterministic scoring
  5. 05RocksDBSync WAL · recovery
Axion architecture showing clients flowing through REST and gRPC APIs, query planning, exact and HNSW indexes, RocksDB storage, and operational telemetry

02 / CAPABILITIES

Infrastructure, not theater.

Every claim below has an implementation and an executable validation path in the repository.

⌁

Native multilayer HNSW

Rust implementation with explicit construction and search breadth controls, stable tie ordering, and exact search as a correctness oracle.

▣

Durable by default

RocksDB is the default engine. Synchronous write batches commit records before acknowledgement and rebuild the derived graph deterministically.

⌘

Metadata-aware

Boolean filters, nested JSON fields, documented exact pre-filter semantics, bounded pagination, and atomic batch upserts.

↔

REST, gRPC, Python

Versioned OpenAPI and protobuf contracts plus typed synchronous and asynchronous Python clients.

◫

Operationally visible

Structured JSON logs, request IDs, Prometheus counters and histograms, optional OTLP tracing, and separate health probes.

⎈

Deployable

Hardened multi-stage images, Compose, Kubernetes resources, secret references, resource limits, health probes, and persistent volumes.

03 / MEASURED SEARCH

Numbers earn their place.

The committed comparison uses the official ANN-Benchmarks framework, its standardized random-xs-20-euclidean dataset, three timed runs, and one pinned CPU.

Recall versus queries per second chart comparing Axion and hnswlib
End-to-end Axion HTTP + RocksDB retrieval versus in-process hnswlib. Full methodology is committed.
RECALL@101.0000ef_search 64
QUERIES / SEC3,873.9ef_search 8 · recall 0.984
DATASET10K × 20D9K train · 1K queries
Inspect raw evidence

Criterion measures Rust components. ANN-Benchmarks measures comparative search behavior. The repository publishes configurations, raw CSV, host metadata, exact revisions, and limitations.

04 / DEPLOYMENT

Persistence is tested through failure.

The kind workflow builds and loads real images, writes authenticated data, deletes the live server pod, waits for replacement, and verifies the vector from the same PVC.

LOCAL

Docker Compose

Non-root read-only containers, a named RocksDB volume, health checks, API authentication, and the full console.

docker compose up --build -d
SCALE-OUT

Partitioned search

Deterministic single-owner sharding, parallel fan-out, stable global merge, and explicit partial-result semantics.

make distributed-test

Clear boundary: the 0.1.0 coordinator partitions but does not replicate data and does not claim consensus, automatic failover, or automatic rebalancing.

05 / OPERATIONS

A real console for a real engine.

These views were captured by Playwright while driving the production Next.js console against the Rust backend. Select a view to inspect it.

axion / console
Axion operational overview

06 / DEVELOPER EXPERIENCE

Useful in four calls.

Run Axion, create a collection, write a vector, and search. The typed client validates locally and preserves structured metadata.

quickstart.py
from axion_client import AxionClient

client = AxionClient("http://localhost:8080")
client.create_collection(
    "documents", dimensions=384, metric="cosine"
)
client.upsert("documents", [{
    "id": "doc-1", "values": embedding,
    "metadata": {"team": "search"}
}])
results = client.search("documents", query, top_k=10)

OPEN SOURCE · MIT LICENSE

Search systems should make their trade-offs visible.

Inspect the engine, reproduce the evidence, and deploy it on infrastructure you control.

Explore Axion on GitHub Contribute