Store: Qdrant

Key Points

Qdrant = open-source vector database. Rust-based; fast; production-ready.
Self-host (Docker / k8s) or Qdrant Cloud (managed).
Strong filter language; payload (metadata) richer than basic vector DBs.
Less native hybrid than Azure AI Search; recent improvements.
When: open-source preference, multi-cloud, Rust-driven perf needs.

Setup

// Qdrant client
var client = new QdrantClient("localhost", 6334);   // gRPC port

var collection = new QdrantCollection<ulong, MyDoc>(client, "my-collection");

await collection.EnsureCollectionExistsAsync();

Record

public class MyDoc
{
    [VectorStoreKey]
    public ulong Id { get; set; }

    [VectorStoreData]
    public string Title { get; set; } = "";

    [VectorStoreData]
    public string Category { get; set; } = "";

    [VectorStoreVector(Dimensions: 1536, DistanceFunction = DistanceFunction.CosineSimilarity)]
    public ReadOnlyMemory<float>? Embedding { get; set; }
}

Search with filter

await foreach (var r in collection.SearchAsync(queryEmb, top: 10, new VectorSearchOptions<MyDoc>
{
    Filter = d => d.Category == "Tech" && d.Title.Contains("dotnet")
}))
{ /* ... */ }

Qdrant filter language is rich: must, must_not, should, range, geo, nested.

Payload

Beyond the typed fields, Qdrant stores arbitrary JSON payload:

public class MyDoc
{
    /* typed fields */
    [VectorStoreData]
    public Dictionary<string, object?>? Payload { get; set; }   // arbitrary
}

HNSW index

Qdrant uses HNSW (Hierarchical Navigable Small World). Tunable m, ef_construct, ef.

new VectorStoreCollectionDefinition
{
    /* ... */
    Properties =
    {
        new VectorStoreVectorPropertyDefinition("Embedding", typeof(ReadOnlyMemory<float>))
        {
            Dimensions = 1536,
            DistanceFunction = DistanceFunction.CosineSimilarity,
            IndexKind = IndexKind.Hnsw
        }
    }
};

Most defaults fine; tune for huge datasets.

Quantization

Qdrant supports scalar, product, binary quantization → reduce memory; minor quality cost.

Hybrid (in 2026)

Qdrant added hybrid (dense + sparse) recently. Combine with BM25-like sparse vectors:

// Sparse vector field via SparseVectors API

Less mature than Azure AI Search hybrid.

Collections / shards

For multi-tenant: multiple collections (one per tenant) OR single collection with tenantId filter. Filter approach scales better.

Self-host

# docker-compose.yml
services:
  qdrant:
    image: qdrant/qdrant:latest
    ports: ["6333:6333", "6334:6334"]
    volumes: ["./qdrant_storage:/qdrant/storage"]

Production: cluster mode for scale + HA.

Qdrant Cloud

Managed offering. Good for teams without ops capacity. Per-region pricing.

Auth

new QdrantClient("host", port, apiKey: "...");

For self-hosted: TLS + API key.

.NET integration

Qdrant.Client SDK + Microsoft.Extensions.VectorData.Qdrant adapter.

Performance

Rust-based; very fast. Production benchmarks: - 10K QPS at sub-10ms p95 on modest hardware. - 100M+ vectors per cluster.