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

data-agent — parity with Microsoft Fabric Data Agent (NL-to-query AI data agent)

Source UI: Microsoft Fabric → Data Science → Data agent. Grounded in Microsoft Learn:

  • https://learn.microsoft.com/fabric/data-science/concept-data-agent
  • https://learn.microsoft.com/fabric/data-science/how-to-create-data-agent
  • https://learn.microsoft.com/fabric/data-science/data-agent-add-datasources
  • https://learn.microsoft.com/fabric/data-science/data-agent-configurations
  • https://learn.microsoft.com/fabric/data-science/data-agent-configuration-best-practices
  • https://learn.microsoft.com/fabric/data-science/data-agent-end-to-end-tutorial
  • https://learn.microsoft.com/fabric/data-science/evaluate-data-agent

Loom surface: apps/fiab-console/lib/editors/phase4/data-agent-editor.tsx → DataAgentEditor (registered for data-agent in lib/editors/registry.ts). Lifecycle pane lib/panes/data-agent.tsx; Config-Copilot lib/editors/data-agent-config-copilot.tsx; result-viz lib/editors/data-agent-result-viz.tsx. Runtime lib/azure/data-agent-client.ts (grounded chat) + lib/azure/data-agent-execute.ts (real read-only per-source query execution). Routes under app/api/items/data-agent/[id]/* (chat / conversations / copilot / evaluate / publish / m365-copilot / deploy).

Last verified: 2026-07-01 against current code. The editor now ships tabs Build · Config Copilot · Test chat · Evaluate · Publish · Consume · Run inspector · Monitoring — Evaluation (#15), Consume (#20), and conversation starters (#22) have shipped and flip ❌→✅ below.

Azure-native by default (no Microsoft Fabric dependency): a data agent is a Cosmos item; grounded chat runs on the Azure OpenAI deployment the cross-item Copilot resolves; sources execute against Synapse (T-SQL), ADLS/Spark (lakehouse), ADX (KQL), Azure AI Search, ADX graph. Publishing to Foundry Agent Service / M365 Copilot is strictly opt-in.

Real feature inventory (every capability, grounded in Learn)

A. Data sources (left "explorer" rail)

  1. Add up to 5 data sources in any combination.
  2. Source types: Lakehouse, Warehouse / SQL DB, KQL database / Eventhouse, Power BI semantic model, Ontology (preview), GQL graph model (preview), Microsoft Graph, Azure AI Search index / unstructured (preview).
  3. Schema selection — a browsable schema tree to check specific Tables / Views / Functions (SQL), Tables / Materialized views / Functions / Shortcuts (KQL), Tables (lakehouse), model tables (semantic). Graph/ontology = whole (no scope).
  4. Data source description — high-level routing context per source.
  5. Data source instructions — table/column descriptions, join logic, value formats, "when asked about X use table Y" routing logic.
  6. Example queries (few-shot) — NL→SQL/KQL/GQL pairs per source; only valid, schema-matching queries are used (Fabric validates and ignores invalid ones); top-3 retrieved by vector similarity at query time.
  7. AI Search unstructured config — Display Name, Search Type (full-text / hybrid / semantic), Number of Documents (3–20), Context/Description, Agent Instructions; citations (URL / file-path fields) returned automatically.
  8. Semantic-model "Prep for AI" — AI Data Schemas, AI Instructions, Verified Answers on the Power BI side; the agent honors them.

B. Agent-level configuration

  1. Agent instructions — up to 15,000 chars, plain-English cross-source routing
  2. term definitions.
  3. Agent name / description — shown to consuming orchestrators for routing.

C. Build / test loop

  1. NL chat test pane (right side) — ask, get grounded answers.
  2. Generated-query transparency — the exact SQL/KQL/DAX run, which source was picked, and (for AI Search) cited documents.
  3. Multi-source routing — agent picks among sources; trace shows which.
  4. Conversation thread with history.

D. Evaluation

  1. Data agent evaluation — supply a ground-truth set of question/expected-answer (or expected-query) pairs, run the agent over all, get an accuracy score + per-question pass/fail, the generated query, and the thread link per row; iterate.

E. Guardrails / governance

  1. Agent answers only from attached sources (scope limiting).
  2. Queries run read-only under the caller's identity (RLS/CLS honored).
  3. Content safety on the model; sensitivity-aware (Purview labels on sources).

F. Publish / consume

  1. Publish a versioned agent; draft vs published version.
  2. Consume programmatically — published agent reachable as a REST endpoint (OpenAI-compatible) + Python SDK; consumable from AI Foundry and Copilot Studio.
  3. Version history — published versions; see/restore current published version.
  4. Conversation starters / suggested prompts surfaced to consumers.

Loom coverage (built ✅ / honest-gate ⚠️ / MISSING ❌)

# Capability Status Notes / backend
1 Up to 5 typed sources ✅ addSource caps at 5; sources are real Loom items via GET /api/items/by-type
2 Source types ⚠️ warehouse / lakehouse / kql / semantic-model / ai-search / ontology / graph built; Microsoft Graph ❌
3 Schema selection ❌ only a comma-separated tables text Input + type-aware label — no schema tree picker (violates no-freeform-config)
4 Data source description ✅ per-source description, fed into the grounded prompt
5 Data source instructions ✅ per-source instructions textarea, seeded with the Fabric template
6 Example queries (few-shot) ⚠️ authored + Config-Copilot-generated; gated correctly per type; no schema validation of examples
7 AI Search unstructured config ❌ ai-search treated as a generic source; no search-type / doc-count / citations
8 Semantic-model Prep-for-AI ⚠️ honest MessageBar points at Power BI Prep for AI
9 Agent instructions (15k) ✅ maxLength=15000 + live counter + template
10 Agent name / description ✅ alias + description
11 NL chat test pane ✅ /chat → live AOAI grounded turn, composer pinned
12 Generated-query transparency ✅ tools-used trace + executed rows (DataAgentResultViz KPI/chart/table toggle)
13 Multi-source routing trace ✅ tools[] per turn
14 Conversation history ✅ /conversations (Cosmos), History menu
15 Evaluation ✅ Evaluate tab: ground-truth Table + Run → POST /evaluate runs each Q through chatGrounded, judges with an AOAI LLM-as-judge (aoaiChatJson), persists evalRuns to Cosmos, shows accuracy donut + per-Q pass/fail + run history
16 Answer only from sources ✅ system prompt scope-limits
17 Read-only under identity ✅ data-agent-execute SELECT/WITH-only, KQL mgmt/ingest blocked, 25-row cap
18 Content safety / sensitivity ⚠️ DSPM labels emitted post-hoc (resolveAgentSourceLabels); no guardrails config UI
19 Publish (Foundry / M365) ✅ / ⚠️ /publish (Foundry Agent Service) + /m365-copilot (Copilot Studio); honest infra-gates
20 Consume / REST endpoint + snippets ✅ Consume tab renders the POST endpoint + copy-paste cURL / Python / JS snippets (lang Dropdown + Copy). Snippets target the existing /chat route (no separate consume/ route / APIM exposure yet)
21 Version history ❌ publishedSnapshot stored but no versions UI / restore
22 Conversation starters ✅ Consume tab authors suggested prompts (add/update/remove), persisted to state.conversationStarters, surfaced in the Test-pane empty state
— Run-steps inspector (published Foundry agent) ✅ extra-vs-Fabric: /data-agent/run-steps trace
— Monitoring (Azure Monitor alert rules) ✅ extra-vs-Fabric: reuses activator rules route

Grade: B+ / A- (functional, real backend). The build/test loop is real and grounded, and Evaluation (LLM-as-judge /evaluate persisting runs to Cosmos), a Consume endpoint + cURL/Python/JS snippets, and conversation starters have now shipped — all three flip ❌→✅. The rows that remain genuinely missing/partial are the real schema-tree picker (#3, still comma-separated text), Microsoft Graph source type (#2), AI Search unstructured config (#7), a Guardrails config surface (#18), and version history (#21) — that's what still reads as "basic."

Build plan (prioritized)

P0 — visible parity uplift

  • Evaluation tab. New Evaluate tab + POST /api/items/data-agent/[id]/evaluate. Author/import a ground-truth set (question + expected answer/query) into a Cosmos data-agent-eval doc; "Run evaluation" iterates each question through chatGrounded, judges with an AOAI LLM-as-judge (correctness + query-match), persists a run with per-question pass/fail + generated query + thread link, shows an aggregate accuracy ring + a results DataGrid. Web5: TileGrid of past runs, EmptyState, status badges.
  • Schema-tree source picker. Replace the comma-separated tables Input with a real checkbox schema browser (Tree/TreeItem of Tables/Views/Functions; KQL adds MVs/ Functions/Shortcuts) fed by a new GET /api/items/data-agent/[id]/source-schema?sourceId= route introspecting the real backend (Synapse INFORMATION_SCHEMA, ADX .show tables/functions, ADLS/Delta catalog, AI Search fields) — much of the introspection already exists in data-agent-execute / Config Copilot. Persist selected objects typed.
  • Microsoft Graph + AI Search source config. Add a microsoft-graph source type (honest-gate on Graph perms) and an AI Search config block (Display Name, Search Type Combobox, Number of Documents SpinButton 3–20, Context). Surface AI Search citations (doc title + link) in the chat trace.

P1

  • Guardrails panel (sub-section/tab). Typed controls (no JSON): max rows returned (SpinButton), allowed/blocked topics (TagPicker), "answer only from sources" Switch, PII/sensitivity redaction Switch (wires to resolveAgentSourceLabels / Purview), AOAI content-safety Switch. Enforced in data-agent-execute (row cap) + chatGrounded (scope/topics) + the chat route.
  • Consume tab. Stable Loom REST endpoint + copy-paste cURL / Python / JS snippets (OpenAI-compatible against the existing /chat route), Entra/API-key auth note, optional "Expose via APIM" action (real APIM import). CodeBlock with copy per language.
  • Example-query validation. On blur / "Validate", dry-run each few-shot example's SQL/KQL against the picked schema via data-agent-execute (parse/EXPLAIN, no data movement); badge each pair valid/invalid and warn invalid pairs are ignored (Fabric behavior).

P2

  • Version history. List publishedSnapshot versions (timestamp + diff vs draft + restore); GET/POST /api/items/data-agent/[id]/versions.
  • Per-agent model + temperature settings (Settings sub-panel) overriding the cross-item AOAI default; Dropdown of deployed models + Slider temperature.
  • Conversation starters authoring — list of suggested prompts persisted to state, emitted on publish (Foundry/M365) + shown in the test-pane empty state.

Bicep sync

P0/P1/P2 extend an existing Cosmos-backed item type and reuse already-wired AOAI (LOOM_AOAI_*), Foundry (LOOM_FOUNDRY_*), Synapse/ADX, AI Search, and Purview env contracts. The eval ground-truth set persists in the existing Cosmos items/agent state (or a data-agent-eval doc in the same container — createIfNotExists); no new Azure resource. APIM exposure (P1 Consume) reuses the existing APIM module if the operator opts in.