[Proposal] dotnet-ai plugin: AI and ML skills for .NET

[luisquintanilla]

Summary

This proposal introduces a new dotnet-ai plugin to the dotnet/skills repository containing a set of discrete, composable AI and ML skills for .NET developers.

The .NET AI/ML ecosystem has grown significantly — Microsoft.Extensions.AI, Microsoft.Agents.AI, ML.NET's deep learning support (TorchSharp), ONNX Runtime, vector data connectors, and the GitHub Copilot SDK all provide powerful capabilities. But developers face a real problem: knowing which library to use when, and how to wire them together correctly. These skills encode that expert knowledge.

Design Philosophy

Each skill follows the agent skills philosophy:

• Purposeful — solves a real, repeatable .NET AI/ML problem
• Concise — encodes decisions, not options
• Scoped — small enough to explain in one or two sentences
• Composable — designed to be reused as part of larger workflows (e.g., RAG pipeline composes embeddings + vector search + ingestion + chat)
• Validated — each skill has an eval.yaml with concrete scenarios

Layering Model

The skills are organized around the .NET AI stack:

Layer	Technology	Skill(s)
Harness	GitHub Copilot SDK	copilot-sdk-integration
Runtime	Microsoft Agent Framework (MAF)	agentic-workflow
Framework	Microsoft.Extensions.AI (MEAI)	meai-chat-integration, meai-embeddings
Classical ML	ML.NET + TorchSharp	mlnet
Inference	ONNX Runtime, Ollama, Foundry Local	onnx-runtime-inference, local-llm-inference
Data	VectorData, DataIngestion	vector-data-search, data-ingestion-pipeline
Composition	RAG = chat + embeddings + vector + ingestion	rag-pipeline
Router	Decision tree across all skills	technology-selection

Proposed Skills

#	Skill	Scope	Reference Files
1	meai-chat-integration	IChatClient setup, middleware pipeline, streaming, structured output, retry/resilience, model pinning, token counting	tokenizers.md
2	meai-embeddings	IEmbeddingGenerator setup, batch generation, dimension selection	—
3	mlnet	Classical ML (classification, regression, clustering, etc.) + deep learning (image classification, object detection, NER, QA, text classification via TorchSharp). Full train, evaluate, deploy lifecycle	custom-transforms.md, dataframe.md, torchsharp.md
4	vector-data-search	Microsoft.Extensions.VectorData abstractions, connector selection, hybrid search	—
5	data-ingestion-pipeline	Microsoft.Extensions.DataIngestion pipeline, chunking strategies, enrichers	—
6	onnx-runtime-inference	ONNX Runtime inference in .NET, standalone and via ML.NET, execution providers (CPU/GPU/DirectML)	tensors.md
7	local-llm-inference	Running LLMs locally via Ollama and Foundry Local through IChatClient	—
8	agentic-workflow	Microsoft Agent Framework: AIAgent, AgentSession, AsAIFunction multi-agent composition, durable agents, A2A protocol	—
9	copilot-sdk-integration	GitHub Copilot SDK three dimensions: LLM backend (zero-config auth, BYOK), agent harness (sessions, MCP, safe outputs), platform extensions. Includes Bridge Pattern (prototype with Copilot, deploy with Azure)	—
10	rag-pipeline	End-to-end retrieval-augmented generation composing ingestion + embeddings + vector search + chat	—
11	technology-selection	Meta/router skill: decision tree that routes developers to the correct skill based on their task	—

Description Size Budget

Per the discussion in #222, description sizes are within proposed limits:

• Largest single description: 692 chars (mlnet)
• Plugin aggregate: 4,918 chars across 11 skills (avg 447)

Relationship to Existing Work

Issue #31 describes the need for .NET Agent Framework authoring guidance. This proposal decomposes that and adjacent AI/ML scenarios into discrete composable skills — agentic-workflow covers the MAF agent authoring piece, while the other skills handle the broader AI/ML landscape.

Plugin Structure

plugins/dotnet-ai/
  plugin.json
  agents/
  skills/
    meai-chat-integration/
      SKILL.md
      references/
        tokenizers.md
    meai-embeddings/SKILL.md
    mlnet/
      SKILL.md
      references/
        custom-transforms.md
        dataframe.md
        torchsharp.md
    vector-data-search/SKILL.md
    data-ingestion-pipeline/SKILL.md
    onnx-runtime-inference/
      SKILL.md
      references/
        tensors.md
    local-llm-inference/SKILL.md
    agentic-workflow/SKILL.md
    copilot-sdk-integration/SKILL.md
    rag-pipeline/SKILL.md
    technology-selection/SKILL.md
tests/dotnet-ai/
  <one eval.yaml per skill>

Suggested Merge Order

Skills have a dependency graph for cross-references. Suggested order:

1. Plugin scaffold — plugin.json, empty dirs, CODEOWNERS, marketplace.json
2. Foundation skills (any order): meai-chat-integration, meai-embeddings, mlnet, vector-data-search + data-ingestion-pipeline, onnx-runtime-inference + local-llm-inference
3. Agent stack: agentic-workflow + copilot-sdk-integration (Bridge Pattern links them)
4. RAG pipeline — composes foundation skills
5. Technology selection — router, references all skills; lands last

Sub-Issues

Each deliverable unit has its own sub-issue with detailed scope, linked below.

• [dotnet-ai] Plugin scaffold #226 Plugin scaffold
• [dotnet-ai] MEAI chat integration #227 MEAI chat integration
• [dotnet-ai] MEAI embeddings #228 MEAI embeddings
• [dotnet-ai] ML.NET training and deployment #229 ML.NET training & deployment
• [dotnet-ai] Vector data search + data ingestion pipeline #230 Vector data search + data ingestion pipeline
• [dotnet-ai] ONNX Runtime + local LLM inference #231 ONNX Runtime + local LLM inference
• [dotnet-ai] Agent stack (agentic-workflow + copilot-sdk-integration) #232 Agent stack (agentic-workflow + copilot-sdk-integration)
• [dotnet-ai] RAG pipeline #233 RAG pipeline
• [dotnet-ai] Technology selection (router) #234 Technology selection (router)

Open Questions

• Does dotnet-ai as a plugin name work, or is there a preferred naming convention?
• Any skills the team wants to defer, cut, or combine?
• Ownership: who should be listed in CODEOWNERS?