Monday, 9:02 a.m. The CFO pings: “What was Q3 gross margin by region—and did audit call out any risks?” Your RAG bot shines on PDFs and wiki pages, but it can’t compute a number you’d put on a KPI card. After Ignite 2025, the answer is cleaner than ever: let a Fabric Data Agent generate and run a governed query for the metric, and let your RAG retriever bring back the one‑sentence risk note. One conversation; two specialized tools; auditable answers.
We’ll start with what changed at Ignite 2025 that matters for builders: Fabric’s new IQ (preview) semantic layer, Ontology as a first‑class data source for agents, deeper Copilot Studio and Foundry Agent Service integrations, and a refreshed SDK. Then we’ll revisit how Data Agents work, contrast them with RAG, and explain why “throwing a table at an LLM” isn’t a substitute. We’ll close with a pragmatic orchestration pattern you can ship.
What changed at Ignite 2025
A semantic layer you can query. Fabric introduced Fabric IQ (preview), a workload that unifies business meaning across OneLake, centered on an Ontology item that models entities/relationships and binds them to data. Fabric IQ is designed to ground copilots and agents in shared concepts, not just schemas. Data Agents now list Ontology alongside Warehouse, Lakehouse, Semantic Models, and KQL as supported sources. Practically, this lets you route “customers with late shipments and cold‑chain breaches last month” to a concept‑aware surface, not a pile of joins.
Graph in Fabric (preview). Fabric IQ’s companion, Graph, adds GQL and NL2GQL over labeled property graphs in OneLake. It’s not a Data Agent source itself, but it underpins ontology traversals and makes relationship-heavy questions first‑class.
Orchestration got real. Azure AI Foundry Agent Service tightened integration with Fabric: add a published Fabric Data Agent as a knowledge/tool in your Foundry agent; orchestration uses on‑behalf‑of identity to run governed NL→SQL/DAX/KQL under the end user. As of the docs, one Foundry agent can attach one Fabric Data Agent. Foundry’s November wave adds managed hosting, built‑in memory, and multi‑agent workflows, which matters when a single turn spans metrics (Fabric) and memos (RAG).
Channels expanded. Copilot Studio can now connect a Fabric Data Agent as a connected agent (preview), so Teams and web copilots can invoke governed, structured queries when the turn calls for numbers.
Tooling matured. The Fabric Data Agent SDK on PyPI added ontology data source support in the Nov 21, 2025 build—useful for evaluation and CI. The Fabric “What’s new” page also highlights a Python evaluation SDK for agents. Treat the agent like code: test, diff, and promote.
So what? You can now ground conversational analytics in shared semantics (IQ), keep unstructured retrieval in a proper RAG lane (Azure AI Search as a Foundry knowledge source), and let the orchestrator pick the right tool per turn—with identity and governance intact.
How Fabric Data Agents work—end to end
A Fabric Data Agent is a governed, read‑only chat interface over structured Fabric sources: Lakehouse/Warehouse (SQL), Power BI semantic models (DAX), KQL databases (KQL), and now Ontology from the Fabric IQ workload. You select which tables (and models) are visible, add instructions and examples (for SQL/KQL), and publish. Users ask in natural language; the agent plans a query, executes it under their identity, and returns an answer with the generated code and the steps it took. Agents are still preview.
Under the hood, the flow is consistent:
Identity & schema. The agent uses the end user’s Entra ID to read schema and execute the query so RLS/CLS and source permissions apply. In Foundry orchestration, on‑behalf‑of identity flows end‑to‑end.
Prompt & tool choice. The agent builds a prompt with the question, live schema/metadata, instructions, and optional examples, then invokes NL2SQL / NL2DAX / NL2KQL (and routes to Ontology where appropriate). Queries are validated before execution.
Execution & transparency. The engine runs the query; the chat shows the intermediate steps and the generated query for auditability and reuse. It remains read‑only by design.
Where it runs: inside Fabric, via a published endpoint to Foundry, or as a connected agent in Copilot Studio. Orchestration models in Foundry don’t replace the models used within Fabric for NL→SQL/DAX/KQL.
How this differs from RAG (and why both belong together)
If your muscle memory is chunk → embed → retrieve → rerank → synthesize, a Data Agent flips the contract.
RAG retrieves, a Data Agent computes. RAG finds passages from unstructured sources through Azure AI Search knowledge bases/sources and synthesizes an answer with citations. A Data Agent emits queries that compute facts in engines built for scale and governance.
Use RAG for policies, memos, tickets, and context; use the agent for KPIs. Reproducibility and lineage differ. RAG quality lives in corpus design and retrieval logs; agents return a re‑runnable query and step trace under user identity. That’s why finance, audit, and ops teams trust the number. Scope of data. Data Agents target structured Fabric artifacts (now including Ontology). Keep unstructured corpora on Azure AI Search as proper knowledge sources.
Best practice: Orchestrate both in Foundry Agent Service. Let the orchestrator choose the Fabric tool when a turn needs governed numbers and the Search knowledge base when it needs narrative or citations—with memory to keep long‑running analyses coherent.
Why “just paste the table into the prompt” isn’t an agent
Streaming a CSV into the LLM feels quick, but it trades away the properties that make assistants operational:
No pushdown; high token cost. You pay to move rows into the model instead of pushing filters/aggregations down to SQL/DAX/KQL over OneLake. Agents plan and push computation to engines.
Governance gaps. Prompt stuffing skirts RLS/CLS and auditing. Agents execute as the end user, inheriting permissions and leaving query lineage.
No reproducible lineage. A pasted table yields no query or steps to re‑run or diff. Agents return both.
If it’s a number you’ll defend, compute it where the data lives. If it’s context you’ll cite, retrieve it from documents.
Orchestrating the Monday‑morning question
Back to the CFO: “margin by region” + “auditor risks.”
Metric path (Data Agent). Route to a Data Agent backed by your finance semantic model (or an ontology concept if margin is defined there). The agent generates NL→DAX under the user’s identity, honors RLS/CLS, and returns a traceable result (code + steps).
Narrative path (RAG). Route to Azure AI Search via a knowledge base for the audit memo. Retrieve passages; synthesize a one‑sentence note with a citation.
Orchestration. Host the assistant in Foundry Agent Service; attach Microsoft Fabric (Data Agent) as a tool and your Azure AI Search knowledge base as another tool. Foundry adjudicates which tool to call; identity flows on‑behalf‑of to Fabric.
Operating like software: evaluation, CI/CD, semantics
Evaluate before you publish. Use the Python evaluation SDK and the fabric‑data‑agent‑sdk to programmatically run ground‑truth suites. The latest SDK build (Nov 21, 2025) adds ontology support, so you can test concept‑level queries too.
Model once, reuse everywhere. Treat IQ’s Ontology as your shared language so copilots, Data Agents, and reports agree on what “customer” and “margin” mean. Use Graph for relationship‑heavy analysis that Ontology enables.
Know the current limits. Data Agents remain preview and read‑only; when used from Foundry, you can attach one Fabric Data Agent per Foundry agent. For channels and multi‑agent bots, Copilot Studio’s connected agent integration is preview.
Summary
Data Agents aren’t “RAG with SQL.” They’re identity‑aware planners that generate and run SQL/DAX/KQL (and can now draw on Ontology in IQ) against governed Fabric sources—returning answers with re‑runnable lineage. RAG remains the right tool for unstructured knowledge with citations. After Ignite 2025, the story is stronger: a shared semantic layer (IQ), deeper Foundry and Copilot Studio integration, and SDKs that let you evaluate and ship like software. Route metrics to engines, context to RAG, and orchestrate both—so Monday morning answers are both trusted and sourced.
