Embeddings

Embedding providers turn text into numeric vectors. The vector is what a vector store indexes, what cosine similarity measures, and what hybrid search combines with BM25.

LogicGrid.Memory serves three providers — local (Ollama) OpenAI Client and OpenAI compatible hosted-API servers (vLLM, TEI, Gemini).

dotnet add package LogicGrid.Memory

EmbeddingClientBase contract

public abstract class EmbeddingClientBase
{
    public abstract string ModelName { get; }
    public abstract int    Dimensions { get; }

    public abstract Task<float[]> EmbedAsync(
        string text, CancellationToken ct = default);

    // Override for providers with native batching (OpenAI, vLLM, TEI).
    // Default implementation calls EmbedAsync in a loop.
    public virtual Task<float[][]> EmbedBatchAsync(
        IList<string> texts, CancellationToken ct = default);
}

Dimensions is sometimes 0 until the first call (Ollama discovers it from the response); pass the known dimension count to the constructor when you have it to avoid that round-trip.

OllamaEmbeddingClient

public OllamaEmbeddingClient(
    string baseUrl       = "http://localhost:11434",
    string defaultModel  = "nomic-embed-text",
    int    dimensions    = 0)   // 0 = discover on first call

Common used models

Model	Dimensions	When to use
nomic-embed-text (default)	768	General-purpose, balanced quality/size.
mxbai-embed-large	1024	Higher quality, slightly slower.
bge-m3	1024	Multilingual.

How to use it

var embedder = new OllamaEmbeddingClient(defaultModel: "nomic-embed-text");
var vec = await embedder.EmbedAsync("Hybrid search combines dense and sparse retrieval.");
Console.WriteLine($"Dimensions: {embedder.Dimensions}");  // 768

OpenAiEmbeddingClient

public OpenAiEmbeddingClient(
    string apiKey,
    string defaultModel = "text-embedding-3-small",
    int    dimensions   = 1536)

The OpenAI API. text-embedding-3-small is the cheap default; text-embedding-3-large is the higher-quality option.

How to use it

var embedder = new OpenAiEmbeddingClient(
    apiKey:       Environment.GetEnvironmentVariable("OPENAI_API_KEY")!,
    defaultModel: "text-embedding-3-small");

OpenAiCompatibleEmbeddingClient

public OpenAiCompatibleEmbeddingClient(
    string  baseUrl,
    string  defaultModel,
    int     dimensions = 0,
    string? apiKey     = null)

Anything that exposes the OpenAI Embeddings API at /v1/embeddings: vLLM, TEI (HuggingFace Text Embeddings Inference), LM Studio, Gemini via its OpenAI-compatible endpoint.

// vLLM serving an embedding model locally
var vllm = new OpenAiCompatibleEmbeddingClient(
    baseUrl:      "http://localhost:8000",
    defaultModel: "BAAI/bge-large-en-v1.5",
    dimensions:   1024);

// TEI
var tei = new OpenAiCompatibleEmbeddingClient(
    baseUrl:      "http://localhost:3000",
    defaultModel: "BAAI/bge-base-en-v1.5",
    dimensions:   768);

Batch embedding

EmbedBatchAsync is significantly cheaper per item — every provider batches under the hood. The RAG pipeline uses it automatically when ingesting documents.

var vectors = await embedder.EmbedBatchAsync(new[]
{
    "First chunk.",
    "Second chunk.",
    "Third chunk.",
});

Where next

• Vector stores — InMemory, Qdrant, Hybrid.
• Semantic memory — durable, role-scoped memory built on top of an embedding model and a vector store.
• RAG pipeline — full end-to-end ingest + search.