Sylin.Koan.Data.Vector

Sylin.Koan.Data.Vector

Supported Entity-first vector persistence and search for Koan. The runtime elects one referenced provider per entity, memoizes repository resolution, applies Koan data isolation to vector operations and physical names, and reports which providers actually became runtime dependencies.

Install

dotnet add package Sylin.Koan.Data.Vector.Connector.InMemory

Provider packages bring this runtime. The in-memory provider is the zero-infrastructure development floor; replace or complement it with a durable provider package when guarantees grow. AddKoan() discovers the runtime and referenced providers; no vector-specific startup code is required. Reference this runtime directly only when authoring a provider or intentionally composing Vector without a provider.

Smallest meaningful use

using Koan.Data.Core.Model;
using Koan.Data.Vector;

public sealed class Article : Entity<Article>
{
    public string Title { get; set; } = "";
}

var article = new Article { Title = "Koan package quality" };
float[] embedding = [0.12f, 0.42f, 0.88f];

await Vector<Article>.SaveWithVector(article, embedding);
var nearest = await Vector<Article>.Search(embedding, topK: 5);

Omitting topK requests 10 matches. Any explicit positive value is preserved through the runtime; adapters do not cap or replace it. Non-positive values fail at VectorQueryOptions construction, before provider I/O.

SaveWithVector persists the entity through the active data provider and its vector through the elected vector provider. Use Vector<Article>.Save(...) when only the vector representation should change.

Decorate an entity only when automatic provider election is not the intended policy:

using Koan.Data.Vector.Abstractions;

[VectorAdapter("qdrant")]
public sealed class Article : Entity<Article>;

The exact provider must be referenced and available; Koan does not silently fall back from an explicit request.

Guarantees and boundaries

• Referencing the runtime without any vector provider leaves Vector<TEntity>.IsAvailable false; operations fail correctively instead of inventing storage.
• A directly referenced provider wins automatic election over the low-priority in-memory floor. Multiple candidates use the shared provider-priority policy; an explicit VectorAdapter request is exact.
• Save writes only the vector store. SaveWithVector coordinates entity and vector persistence; outside a Koan data transaction, an entity success followed by vector failure raises VectorCoordinationException and requires re-embedding/reconciliation.
• Filter, hybrid search, continuation, export, flush, and embedding retrieval depend on selected-provider capability.
• The in-memory provider is process-local and ephemeral. It is not a durability, distribution, or production-scale similarity guarantee.
• Vector dimension, distance metric, score interpretation, consistency, and indexing latency remain provider/model responsibilities.

See TECHNICAL.md for election, caching, naming, isolation, and health behavior.