Skip to content
CSA Loom — the Microsoft Fabric experience for Azure tenants where Fabric isn't yet available: lakehouses, warehouses, notebooks, semantic models, Activator rules, Data Agents, across Commercial, GCC, GCC-High, and DoD IL5

Tutorial 05 — Data Agent over Lakehouse

Author, test, and publish a Loom Data Agent that answers natural-language questions about your Silver table from Tutorial 02. 30 minutes.

Prerequisites

  • Workspace with noaa_silver_daily table from Tutorial 02
  • (Optional) The Power BI model from Tutorial 03
  • An AOAI deployment (e.g. gpt-4o) in your boundary's region

How the Data Agent editor works

A Data Agent is an item with three tabs — Build, Test chat, and Publish. You configure sources + instructions on Build, try it on Test chat (grounded against the real backends), and Publish to Foundry Agent Service. There is no per-workspace "Data Agents pane" and no public chat endpoint for outside consumers — the Test chat tab calls POST /api/items/data-agent/<id>/chat, and publishing registers the agent with Foundry.

Steps

1. Create a Data Agent

Left nav → Data agents (the global page) or open your workspace → New item → category Fabric IQ → Data agent → Create. The editor opens at /items/data-agent/<id>.

2. Build — instructions

On the Build tab, fill the Instructions textarea (the system prompt). Include any few-shot Q→query examples and field notes directly in this text block — there is no separate examples grid:

You are an expert at answering questions about NOAA daily weather data.
Always cite the underlying query you generated. Prefer SQL over the
lakehouse for detail questions; prefer DAX over the semantic model for
aggregate questions like "monthly average temperature".

Column notes:
- temperature_c: Temperature in Celsius, converted from Fahrenheit
- date: Date of observation (UTC)
- station_id: NOAA station identifier (e.g. GHCND:USW00094728)

Examples:
Q: What was the average temperature in January?
SQL: SELECT AVG(temperature_c) FROM noaa_silver_daily WHERE MONTH(date) = 1

Q: How many days did we record above 100F?
SQL: SELECT COUNT(*) FROM noaa_silver_daily WHERE temperature_c > 37.7

Q: Monthly average temperature year-over-year
DAX: EVALUATE SUMMARIZE(noaa_silver_daily, YEAR(date), MONTH(date),
     "AvgTemp", AVERAGE(noaa_silver_daily[temperature_c]))

3. Build — sources

Still on Build, add up to 5 sources. For each: pick a type (Warehouse, Lakehouse, KQL database, Semantic model, AI Search, Ontology, or Graph model), then pick the item from the picker (loaded from GET /api/items/by-type?types=<itemType>). Each source card lets you scope it to specific tables (comma-separated) and add per-source instructions.

For this tutorial add:

  1. Lakehouse → noaa_silver_daily
  2. (Optional) Semantic model → the Power BI model from Tutorial 03

4. Test chat

Switch to the Test chat tab. Ask questions in the composer and click Send. Each turn POSTs { question, history } to POST /api/items/data-agent/<id>/chat. Try:

  • "What was the average temperature in January?"
  • "How many days were above 100°F?"
  • "Show me the trend over the year"

Each response shows the natural-language answer, the generated query (SQL/KQL/DAX), the sourceUsed, and the executed tools (with rowCount, columns, and rows — or an honest gate if a backend isn't provisioned). Tabular tool results render inline via DataAgentResultViz.

5. Publish

Switch to the Publish tab. Enter a description and an optional alias, then click Publish (POSTs { description, alias } to POST /api/items/data-agent/<id>/publish).

  • Commercial (Foundry available): publishResult.ok = true and the tab shows the Foundry agent artifact ID and publishedAt. Use the Inspect form to look the agent up by artifact ID. Consumers call the published Foundry agent.
  • Gov (no Foundry): publishResult.deferred = true with an honest banner "Foundry Agent Service not configured". Use the Test chat tab to interact with the agent until an MCP-compatible registration path is available in your boundary.

What's next

Cleanup

  • Delete the Data Agent from the workspace item tree (right-click → Delete), or leave it — agents are cheap when idle

Troubleshooting

  • Agent gives the wrong query: add more examples and tighten the instructions
  • Agent doesn't answer: check AOAI throttling per Copilot throttling runbook
  • A source returns a gate: provision the named backend (env var / role) it reports