The retrieval engine

Northline is a vector database written for filtered, hybrid, real-time retrieval. The open engine and Cloud speak the same API.

What is different

• Filters during the walk

  Payload predicates are applied while the HNSW graph is traversed. A tenant filter does not mean “fetch 200, keep 5”.

• Hybrid in one query

  Dense similarity and sparse keyword scores fuse before the response. You do not maintain a second sparse index for ordinary BM25-style terms.

• Quantization

  Scalar and binary quantization shrink resident memory. Rescore the shortlist at full precision when the last mile matters.

• Multitenancy

  Isolate customers with a payload key and a scoped API key, or with a collection per tenant when the boundary must be physical.

A first query

from northline import Client

db = Client("http://localhost:6333")

db.open_collection(
    "places",
    size=768,
    distance="cosine",
)

db.put_points(
    "places",
    points=[{
        "id": "lx-1",
        "vector": embedding,
        "payload": {"city": "Lisbon", "rating": 4.8},
    }],
)

hits = db.search(
    "places",
    vector=query_embedding,
    filter={"city": "Lisbon", "rating": {">=": 4}},
    limit=8,
)

Where it runs

• Managed Cloud

  A dedicated cluster on your chosen region. Backups, scaling, and the web console are included.

• Hybrid Cloud

  Data stays in your network. The control plane is managed, the data plane runs on your Kubernetes.

• Private Cloud

  Full isolation for air-gapped and regulated estates, with a custom uptime agreement.

• Edge

  A smaller build for stores, vehicles, and factories that cannot wait on a region.