CanDoItAll.AgentFramework.SemanticCompletion.Onnx 0.1.18

CanDoItAll Agent Framework Semantic Completion — ONNX

Optional local ONNX Runtime embeddings for CanDoItAll.AgentFramework.SemanticCompletion.Driver, targeting .NET 10. This package contains the ONNX adapter, BERT WordPiece tokenizer, and explicit Hugging Face model acquisition helper. It depends on the Driver contracts; the Driver does not depend on this package.

No model binaries are bundled.

Install

dotnet add package CanDoItAll.AgentFramework.SemanticCompletion.Onnx

Create and own a provider directly

using CanDoItAll.AgentFramework.SemanticCompletion.Onnx;

using var embeddings = new OnnxAgentTextEmbeddingGenerator(
    new OnnxAgentTextEmbeddingOptions
    {
        ModelDirectory = ".models/Xenova/all-MiniLM-L6-v2",
        ModelFileName = "onnx/model.onnx",
        VocabularyFileName = "vocab.txt",
        MaxTokenCount = 128,
        NormalizeOutput = true,
    });

OnnxAgentTextEmbeddingGenerator owns an ONNX InferenceSession and implements IDisposable. Reuse one instance rather than creating a session per request. Dispose directly constructed instances. When the DI container constructs the singleton, dispose the service provider/host so the session is released.

Use with the Driver DI extension

Register ONNX before the Driver. The Driver uses TryAdd, so the custom provider remains selected:

using CanDoItAll.AgentFramework.SemanticCompletion.Driver.DependencyInjection;
using CanDoItAll.AgentFramework.SemanticCompletion.Driver.Embeddings;
using CanDoItAll.AgentFramework.SemanticCompletion.Onnx;
using Microsoft.Extensions.DependencyInjection;

services.Configure<OnnxAgentTextEmbeddingOptions>(options =>
{
    options.ModelDirectory = ".models/Xenova/all-MiniLM-L6-v2";
    options.MaxTokenCount = 128;
});
services.AddSingleton<IAgentTextEmbeddingGenerator, OnnxAgentTextEmbeddingGenerator>();
services.AddAgentSemanticCompletionDriver();

The singleton can receive concurrent GenerateAsync calls after the Driver's one-time intent initialization. Validate concurrency and native-runtime settings for the host's workload. The Driver checks that every candidate and query uses the same profile ID and dimension, and always computes bounded full cosine similarity. NormalizeOutput defaults to true, but correctness does not rely on that declaration.

Acquire model assets explicitly

The defaults expect the Hugging Face repository Xenova/all-MiniLM-L6-v2, files onnx/model.onnx and vocab.txt, under .models/Xenova/all-MiniLM-L6-v2.

using var httpClient = new HttpClient();
var downloader = new HuggingFaceModelDownloader(httpClient);
var token = Environment.GetEnvironmentVariable("HF_TOKEN")
    ?? Environment.GetEnvironmentVariable("HUGGINGFACE_TOKEN")
    ?? Environment.GetEnvironmentVariable("HUGGING_FACE_HUB_TOKEN");

await downloader.DownloadAsync(
    new HuggingFaceModelDownloadOptions
    {
        RepositoryId = "Xenova/all-MiniLM-L6-v2",
        Revision = "main",
        OutputDirectory = ".models/Xenova/all-MiniLM-L6-v2",
    },
    token);

Keep access tokens in the host environment or secret store; do not put them in source, options files, logs, or package content. The downloader sends a bearer token only when one is supplied, creates missing directories, and reuses existing non-empty files. Pin a reviewed model revision for reproducible production deployments.

Licensing and trust

This NuGet package does not redistribute Hugging Face model weights. Model selection, download authorization, integrity/provenance checks, storage, and compliance with the selected model's license are the consuming application's responsibility. Microsoft ONNX Runtime and other transitive dependencies retain their own licenses. Review the package dependency list and the model repository before distribution.

Neural similarity remains fallback evidence; it is not authoritative workflow completion, authorization, or proof.

Project links

• Repository
• Driver package
• Architecture and publishing boundaries
• License
• Support and issues
• Security policy
• CanDoItAll