Lodestar.Extensions.AI 0.1.0. This page is frozen at that release. Read the current documentation for what main says now. A link to a decision or a migration page follows main, and leaves the archive.

Presents an OnnxTextEmbedder as an IEmbeddingGenerator<string, Embedding<float>>.

public sealed class OnnxEmbeddingGenerator : IEmbeddingGenerator<string, Embedding<float>>

Constructor — takes the embedder to adapt, the BatchEncoder it should tokenize with, and optionally a model identifier to report. It takes ownership of the embedder: disposing the generator disposes the session.

Example — the shape of a call. It is not executed: see below.

using Lodestar.Embeddings.Persistence;
using Lodestar.Embeddings.Tokenization;
using Lodestar.Extensions.AI;
using Lodestar.Onnx;
using Microsoft.Extensions.AI;

var tokenizer = new WordPieceTokenizer(VocabTxtLoader.Load("vocab.txt"));

using IEmbeddingGenerator<string, Embedding<float>> generator = new OnnxEmbeddingGenerator(
    new OnnxTextEmbedder("model.onnx", tokenizer), new BatchEncoder(tokenizer), "my-model");

var about = generator.GetService(typeof(EmbeddingGeneratorMetadata)) as EmbeddingGeneratorMetadata;

Remarks — every fence on this page and its members is docs-run: skip, and that is not an oversight. A running example would need a model of tens of megabytes, and weights are never committed to this repository — decisions/0003 is the rule, and the packaging sample declares the same exclusion for the same reason. The fences are still compiled against the packed package, so a renamed member still fails CI.

The work is synchronous. The interface is asynchronous because most implementations of it call a service over a network; this one runs a model in the calling process, so OnnxEmbeddingGenerator.GenerateAsync does its work on the caller's thread and hands back an already-completed task. Wrapping it in Task.Run would move the same CPU to a pool thread and tell the caller nothing true.

Ownership is taken, not shared. IEmbeddingGenerator is IDisposable, and a consumer holding this through the interface has no way to learn that disposing it would leave a native session open. So OnnxEmbeddingGenerator.Dispose disposes the embedder. A caller who wants to keep the embedder alive builds a second one.

Applies to — net10.0, netstandard2.0.

See also — the ONNX inference reference, the semantic search guide, the Python equivalence table.