ExpandVectorStore.Qdrant 1.0.4

ExpandVectorStore.Qdrant

Qdrant vector store provider for Microsoft.Extensions.VectorData.

This package lets you use Qdrant through the standard VectorStore and VectorStoreCollection<TKey, TRecord> abstractions. It is intended for RAG, semantic search, and applications that already use Microsoft.Extensions.AI embeddings.

Features

• Create, delete, list, and open Qdrant collections.
• Validate existing collection vector size and distance function against the record definition.
• Upsert, retrieve, delete, scroll, and vector search records.
• Translate common LINQ filters to Qdrant filters for scroll and vector search.
• Delete records by translated Qdrant filters without reading ids first.
• Read all matching records through Qdrant cursor-based scroll.
• Map records from VectorStoreKey, VectorStoreData, and VectorStoreVector attributes.
• Use dynamic dictionary records with VectorStoreCollectionDefinition.
• Create and delete payload indexes for data properties marked as indexed or full-text indexed.
• Serialize common payload values, including strings, numbers, booleans, Guid, DateTime, DateTimeOffset, DateOnly, arrays, and lists.
• Use ReadOnlyMemory<float>, float[], or Embedding<float> for vectors.

Install

dotnet add package ExpandVectorStore.Qdrant

Quick start

using ExpandVectorStore.Qdrant;
using Microsoft.Extensions.VectorData;

var store = new QdrantVectorStore("localhost");
var collection = store.GetCollection<ulong, Product>("products");

await collection.EnsureCollectionExistsAsync();

await collection.UpsertAsync(new Product
{
    Id = 1,
    Name = "Notebook",
    Description = "A compact notebook for meeting notes.",
    Vector = embedding.Vector
});

await foreach (var result in collection.SearchAsync(queryEmbedding.Vector, top: 3))
{
    Console.WriteLine($"{result.Record.Name}: {result.Score}");
}

string[] ids = ["1", "2", "3"];
Product[] selected = await collection
    .GetAsync(product => ids.Contains(product.DataId), top: 100)
    .ToArrayAsync();

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

    [VectorStoreData]
    public string DataId { get; set; } = string.Empty;

    [VectorStoreData(IsIndexed = true)]
    public string Name { get; set; } = string.Empty;

    [VectorStoreData(IsFullTextIndexed = true)]
    public string Description { get; set; } = string.Empty;

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

LINQ filters

The provider translates common expressions into Qdrant filters:

• Boolean logic: &&, ||, and !.
• Comparisons: ==, !=, >, >=, <, and <=.
• Batch membership: ids.Contains(record.DataId).
• Point id membership when the filtered property is [VectorStoreKey].
• Text matching: record.Description.Contains("notebook").
• Null checks: record.OptionalValue == null and record.OptionalValue != null.

The same filter translation is used by GetAsync(filter, top, ...), SearchAsync(vector, top, new VectorSearchOptions<TRecord> { Filter = ... }), and Qdrant-specific filter delete.

Qdrant-specific operations

The standard VectorStoreCollection<TKey, TRecord> abstraction does not expose every Qdrant operation. Use GetService to access the provider-specific interface when you need collection validation, payload index lifecycle, full cursor scroll, or native filter delete:

var qdrantCollection =
    (IQdrantVectorStoreCollection<ulong, Product>)collection.GetService(
        typeof(IQdrantVectorStoreCollection<ulong, Product>))!;

await qdrantCollection.ValidateCollectionAsync();
await qdrantCollection.EnsurePayloadIndexesAsync();

await foreach (var product in qdrantCollection.ScrollAsync(
    product => product.Description.Contains("notebook"),
    new QdrantScrollOptions { BatchSize = 512 }))
{
    Console.WriteLine(product.Name);
}

await qdrantCollection.DeleteAsync(product => product.DataId == "obsolete");

EnsureCollectionExistsAsync() also validates an existing collection before returning. If the Qdrant collection already exists with a different vector size, distance function, or managed payload index type, the provider throws instead of silently using the wrong vector space. Missing payload indexes are created automatically for [VectorStoreData(IsIndexed = true)] and [VectorStoreData(IsFullTextIndexed = true)] properties.

Connection options

Use an existing Qdrant.Client.QdrantClient when your application owns client configuration:

using ExpandVectorStore.Qdrant;
using Qdrant.Client;

var qdrantClient = new QdrantClient("localhost", port: 6334);
var store = new QdrantVectorStore(qdrantClient);

Or let the store create and own the client:

var store = new QdrantVectorStore(
    host: "localhost",
    port: 6334,
    https: false,
    apiKey: "<api-key>");

Current limitations

• One unnamed vector property per record is supported.
• OrderBy expression translation is not implemented yet.
• Filters cannot be applied to vector properties.
• Qdrant point keys support Guid, GUID strings, and non-negative integer key types.