Skip to content

README.md

Latest commit

 

History

History
66 lines (47 loc) · 2.29 KB

README.md

File metadata and controls

66 lines (47 loc) · 2.29 KB

Fabric Data Agent Plugin

Create, test, and tune Microsoft Fabric Data Agents from VS Code using natural language.

What It Does

This plugin provides an agent and three skills for managing Fabric Data Agents through GitHub Copilot:

  • Fabric Data Agent Manager (agent) — Full lifecycle: create → configure → publish → query → tune
  • #fabric-data-agent-create (skill) — Guided end-to-end agent setup with SQL validation
  • #fabric-data-agent-test (skill) — CSV-based accuracy testing with tolerance matching
  • #fabric-data-agent-tune (skill) — Diagnose and fix failing queries

Prerequisites

  • Azure CLI (az login) for Fabric API authentication
  • uv — Python package runner (for the MCP server)
  • Fabric workspace access (Contributor role)

Setup — Connect the MCP Server

This plugin requires the Fabric Data Agent MCP server to provide the tools Copilot uses.

Add this to your .vscode/mcp.json:

{
  "servers": {
    "fabric-data-agent": {
      "command": "uvx",
      "args": [
        "--from", "git+https://github.com/harigouthami/fabric-copilot-plugins.git#subdirectory=fabric-data-agent-mcp",
        "fabric-data-agent-mcp"
      ]
    }
  }
}

Then reload VS Code (Ctrl+Shift+P → "Reload Window") — the 20 MCP tools will appear in Copilot Chat.

No cloning or manual installation needed. uvx automatically fetches, installs, and runs the MCP server.

Example Usage

You: Create a data agent called ADOWIA in A3PInsights workspace
Copilot: ✅ Created. Which lakehouse to connect?

You: External
Copilot: ✅ Connected. Found 4 schemas, 64 tables. Which tables?

You: The tca_adowia* tables from TCA schema
Copilot: ✅ 7 tables selected and verified.

You: [pastes Git repo URL with semantic model]
Copilot: [generates instructions from TMDL files, validates SQL, adds few-shots]
         ✅ Published. Testing: "total time saved" → 7,496.5 hours

Key Features

  • SQL validation: Every few-shot query is tested against the database before adding
  • Knowledge from Git: Fetches TMDL files from ADO repos to auto-generate instructions
  • Accuracy testing: CSV-based test runner with configurable tolerance
  • Tune loop: Reproduce → Diagnose → Fix → Publish → Re-test in one conversation