IMentorQueryEmbedding interface

Model routing, semantic tool filtering and memory relevance already share it. A host's IMentorRagSource is the fourth consumer of the same sentence and the only one outside the library, so it is offered here: ask for the vector instead of embedding the query again, and the turn makes one round trip instead of two.

public sealed class MyRagSource(IMentorQueryEmbedding query, IEmbeddingGenerator<string, Embedding<float>> embeddings)
    : IMentorRagSource
{
    public async Task<IReadOnlyList<MentorRagResult>> SearchAsync(string q, int maxResults, CancellationToken ct = default)
    {
        var vector = await query.GetAsync(q, embeddings, ct);   // free if something already embedded it
        return Search(vector, maxResults);
    }
}

Scoped: the vector belongs to the session that asked for it and is replaced by the next message. Registered by AddMentorAgent(); the generator you pass is the one used on a cache miss.

Namespace: Novior.MentorAgent.Rag. Package: Novior.MentorAgent.

public interface IMentorQueryEmbedding

dotnet add package Novior.MentorAgent

Task<ReadOnlyMemory<float>> GetAsync(string text, IEmbeddingGenerator<string, Embedding<float>> generator, CancellationToken ct = default)

Returns the vector for text, generating it only if it is not the one already computed for this turn. An empty text returns an empty vector without calling out.