[dotnet-ai] Technology selection (router)

plugins/dotnet-ai/skills/technology-selection/SKILL.md

@@ -0,0 +1,149 @@
+---
+name: technology-selection
+description: >
+  META/ROUTER skill — routes developers to the correct .NET AI/ML skill
+  based on their task type. Use this when you are unsure which .NET AI
+  technology or library to apply. Do NOT use this for projects targeting
+  .NET Framework (requires .NET 10+), pure data engineering/ETL with no
+  ML/AI component, or custom deep learning training loops beyond what
+  ML.NET and TorchSharp provide.
+---
+
+# Technology Selection
+
+Route to the right .NET AI/ML skill. Answer the questions below; follow the first matching rule.
+
+## Inputs
+
+| Input | Required | Description |
+|---|---|---|
+| `task_description` | yes | What the developer wants to build or solve |
+| `data_type` | no | structured/tabular, text, image, audio, mixed |
+| `deployment_target` | no | cloud, edge, local/offline |
+| `target_framework` | no | .NET version (must be .NET 10+) |
+
+## Library Stack (reference only)
+
+- **Abstraction** — Microsoft.Extensions.AI (MEAI): always the foundation for LLM work.
+- **Provider SDK** — OpenAI, Azure.AI.OpenAI, OllamaSharp, GitHub.Copilot.SDK, etc.
+- **Runtime** — Microsoft Agent Framework (Microsoft.Agents.AI, prerelease): agent lifecycle, sessions, multi-agent.
+- **Harness** — GitHub Copilot SDK: batteries-included agent runtime with session persistence, MCP, safe outputs.
+
+## Decision Tree
+
+### 1 — Structured Data / Classical ML?
+
+IF the task involves classification, regression, clustering, anomaly detection,
+recommendation, or time-series forecasting on structured/tabular data
+THEN invoke `/mlnet`.
+
+> ⚠️ Do NOT use an LLM for tasks ML.NET handles well (classification on structured
+> data, regression, clustering). LLMs are slower and more expensive for these tasks.
+
+### 1b — Image Classification, Object Detection, NER, QA, or Text Classification?
+
+IF the task involves image classification, object detection, named entity recognition,
+question answering, text classification, or sentence similarity AND the goal is to
+train or fine-tune a model in .NET (not just call an LLM API)
+THEN invoke `/mlnet` (deep learning tasks via TorchSharp).
+
+> Use ML.NET when you need a trained model you own and deploy — not a hosted LLM API call.
+
+### 2 — Natural Language Generation / Chat / Reasoning?
+
+IF the task involves text generation, summarization, reasoning, or chat
+THEN invoke `/meai-chat-integration`.
+
+### 3 — Text Embeddings?
+
+IF the task involves semantic similarity or producing vector representations of text
+THEN invoke `/meai-embeddings`.
+
+### 4 — Vector Storage & Search?
+
+IF the task involves storing, indexing, or querying vector embeddings
+THEN invoke `/vector-data-search`.
+
+### 5 — Document Ingestion for RAG?
+
+IF the task involves document loading, chunking, or enrichment to feed a retrieval pipeline
+THEN invoke `/data-ingestion-pipeline`.
+
+### 6 — End-to-End RAG Pipeline?
+
+IF the task requires retrieval-augmented generation combining ingestion, embeddings,
+vector search, and chat completion
+THEN invoke `/rag-pipeline` (composes `/meai-chat-integration` + `/meai-embeddings`
++ `/vector-data-search` + `/data-ingestion-pipeline`).
+
+### 7 — Multi-Step Agent Workflows?
+
+IF the task involves autonomous multi-step reasoning, tool calling, agent loops,
+multi-agent orchestration, or durable agent sessions
+THEN invoke `/agentic-workflow` (Microsoft Agent Framework — AIAgent, AgentSession, AsAIFunction).
+
+For production scenarios needing persistence, horizontal scaling, or human-in-the-loop,
+`/agentic-workflow` covers durable agents backed by external storage.
+
+### 8 — Pre-Trained ONNX Model Inference?
+
+IF the task involves running a pre-trained ONNX model for inference
+THEN invoke `/onnx-runtime-inference`.
+
+### 9 — Local / Offline LLM Inference?
+
+IF the deployment target is local or offline and the task needs an LLM
+THEN invoke `/local-llm-inference` (Ollama or Foundry Local).
+
+### 10 — GitHub Copilot SDK?
+
+IF the task involves any of these:
+- Building a GitHub Copilot extension for IDE/CLI
+- Using CopilotClient as a zero-config LLM backend (no API key management)
+- Leveraging Copilot harness features (session persistence, MCP servers, safe outputs)
+- Prototyping agents quickly with `gh` CLI auth
+THEN invoke `/copilot-sdk-integration`.
+
+> ⚠️ If the task requires Entra ID / managed identity auth, horizontal scaling,
+> or custom agent loop control — use `/agentic-workflow` with Azure.AI.OpenAI instead.
+> See the Bridge Pattern in `/copilot-sdk-integration` for the prototype-to-production path.
+
+### 11 — Zero-Config Prototyping / No API Keys?
+
+IF the developer wants to get started quickly without managing API keys
+and has a GitHub Copilot subscription
+THEN invoke `/copilot-sdk-integration` (Path A: CopilotClient as LLM Backend).
+
+### 12 — Prototype-to-Production Bridge?
+
+IF the developer wants to prototype with Copilot and later deploy to Azure
+THEN invoke `/agentic-workflow` + `/copilot-sdk-integration`.
+Use CopilotClient during development, swap to Azure.AI.OpenAI for production.
+The AIAgent abstraction keeps agent code unchanged across backends.
+
+### 13 — Hybrid: Structured ML + Natural Language?
+
+IF the task requires BOTH structured ML predictions AND natural language capabilities
+THEN invoke `/mlnet` + `/meai-chat-integration`.
+
+### Anti-Patterns
+
+- **Using `Microsoft.SemanticKernel` for new projects** → Use `Microsoft.Extensions.AI` + `Microsoft.Agents.AI` instead. Semantic Kernel is superseded by these newer abstractions for LLM orchestration and tool calling.
+- **Using Copilot SDK for production enterprise apps requiring Entra ID** → Copilot SDK BYOK only supports static credentials. Use MAF + `Azure.AI.OpenAI` with `DefaultAzureCredential` for enterprise Azure deployments.
+- **Assuming Copilot SDK scales horizontally** → Copilot session state is per-machine (local filesystem). For multi-instance deployments, use MAF durable agents with external storage.
+- **Using an LLM for structured data tasks** → Use ML.NET. LLMs are slower and more expensive for classification, regression, and clustering on structured data.
+
+## Cross-References
+
+| Skill | Invoke |
+|---|---|
+| ML.NET (classical ML) | `/mlnet` |
+| MEAI Chat Integration | `/meai-chat-integration` |
+| MEAI Embeddings | `/meai-embeddings` |
+| Vector Data Search | `/vector-data-search` |
+| Data Ingestion Pipeline | `/data-ingestion-pipeline` |
+| RAG Pipeline | `/rag-pipeline` |
+| Agentic Workflow (MAF) | `/agentic-workflow` |
+| ONNX Runtime Inference | `/onnx-runtime-inference` |
+| Local LLM Inference | `/local-llm-inference` |
+| Copilot SDK Integration | `/copilot-sdk-integration` |

tests/dotnet-ai/technology-selection/eval.yaml

@@ -0,0 +1,58 @@
+scenarios:
+  - name: "Route tabular classification to ML.NET"
+    prompt: "I have a CSV file with customer data (age, income, purchase history) and I want to predict whether a customer will churn. What .NET library should I use?"
+    assertions:
+      - type: "output_contains"
+        value: "ML.NET"
+      - type: "output_not_contains"
+        value: "Microsoft.Extensions.AI"
+    rubric:
+      - "Correctly identifies this as a tabular classification task"
+      - "Routes to ML.NET, not to an LLM-based solution"
+      - "Mentions that LLMs are not appropriate for tabular classification"
+    timeout: 120
+
+  - name: "Route chat integration to MEAI"
+    prompt: "I want to add a chatbot to my ASP.NET Core API that can summarize documents and answer questions in natural language."
+    assertions:
+      - type: "output_contains"
+        value: "Microsoft.Extensions.AI"
+    rubric:
+      - "Identifies this as a natural language task requiring an LLM"
+      - "Routes to Microsoft.Extensions.AI / IChatClient"
+      - "Does not suggest ML.NET for this task"
+    timeout: 120
+
+  - name: "Route RAG to rag-pipeline"
+    prompt: "I need to build a system where users can ask questions about our internal documentation PDFs and get grounded answers."
+    assertions:
+      - type: "output_contains"
+        value: "RAG"
+    rubric:
+      - "Identifies this as an end-to-end RAG scenario"
+      - "Mentions document ingestion, vector search, and LLM grounding"
+      - "Routes to the RAG pipeline approach, not just a simple chat"
+    timeout: 120
+
+  - name: "Route zero-config prototyping to Copilot SDK"
+    prompt: "I want to quickly prototype an AI agent in .NET without managing any API keys. I have a GitHub Copilot subscription."
+    assertions:
+      - type: "output_contains"
+        value: "Copilot"
+    rubric:
+      - "Routes to Copilot SDK for zero-config prototyping"
+      - "Mentions CopilotClient as an LLM backend with gh CLI auth"
+      - "Does NOT present Copilot SDK as only for platform extensions"
+    timeout: 120
+
+  - name: "Route enterprise agent to MAF, not Copilot SDK"
+    prompt: "I need to build an AI agent for our Azure-hosted platform that uses Entra ID for authentication and scales across Kubernetes pods. Should I use the Copilot SDK?"
+    assertions:
+      - type: "output_contains"
+        value: "Microsoft.Agents.AI"
+    rubric:
+      - "Routes to MAF (agentic-workflow) for this enterprise scenario"
+      - "Explains Copilot SDK limitations: no Entra ID, per-machine sessions"
+      - "Recommends Azure.AI.OpenAI with DefaultAzureCredential"
+      - "May suggest the Bridge Pattern for initial prototyping"
+    timeout: 120