Skip to main content
Flux Intelligence is the umbrella capability that brings AI-assisted investigation to your Azure FinOps workspace. It has three components: Ask Flux, the read-only conversational assistant you interact with; the Flux Signals engine, the deterministic optimization rule set that runs independently of the assistant; and the governed tool catalog, the 19 bounded server-side APIs that power both experiences. Every number in an Ask Flux answer came from a named governed tool — the tools are listed in each reply so you can trace any figure back to its source.

​
Governed tool model

Most “AI for your data” approaches hand a model a database connection and rely on generated SQL being correct. Flux does the opposite. The model has no database connection, no Azure credential, no Rill endpoint, and no arbitrary query interface. It can invoke only 19 declared server-side tools, each of which validates and bounds its arguments before calling the same governed services the Flux UI uses. The practical consequences of this design:
  • Reproducibility: any number in an answer came from a named tool over governed data.
  • Honest gaps: if a subscription’s cost export is missing, the answer says so before stating a total. Coverage gaps are stated, not papered over.
  • Bounded scope: the model cannot be prompted into reading something it was never granted — there is no query surface to redirect.
FLUX_INTELLIGENCE_AI_ENABLED defaults to false. AI features require explicit opt-in by setting this variable to true plus supplying credentials for a configured provider.

​
Tool catalog

Tools are bounded to 50 results per search call. FLUX_AI_MAX_TOOL_CALLS (default 12) limits how many tool invocations one answer may make.

​
Cost and billing

​
Optimization and right-sizing

​
Inventory, telemetry, and governance

​
Reference

​
Mutation boundary

Ask Flux has no cloud mutation capability. It cannot start, stop, resize, delete, tag, or purchase anything in Azure, and it cannot write to the analytical store. One tool does create Flux application state: create_rightsizing_board creates a new, empty planning board as a scratch space for a scenario (e.g. “Aggressive downsize option”). Its constraints are deliberate:
  • The new board is never primary and never affects the fiscal outlook.
  • The tool may only be called after the user has explicitly confirmed the exact board name in a later message.
  • Existing boards, placements, and decisions remain human-owned — the assistant cannot alter them.

​
User experiences

Ask Flux panel

Opens as a right-side panel from any authenticated Flux page. Shares the same in-memory conversation until the page is refreshed or the user clears it.

Intelligence Workspace

Full-page investigation experience. Shows 30-day average and p95 browser-to-render duration, quality review summary, and per-answer stage bottleneck details.
Both experiences share the same governed tool catalog, the same authorization model, and the same quality scoring.

​
Analysis profiles

The model service is hidden behind a provider adapter and configured through secure environment settings. The UI and governed tools are not coupled to any named model.

​
Reply structure

A validated Ask Flux reply is a structured document, not free text. It can carry:
  • Summary — a concise headline answer
  • Markdown — formatted analysis and tables (raw HTML is not rendered)
  • Recharts specifications — governed chart definitions (line, bar, or area; bounded rows and series)
  • Mermaid diagrams — strict security mode; click directives, custom classes, and HTML are rejected
  • Retrieved facts — data extracted verbatim from tool results, kept separate from interpretation
  • Interpretation — the model’s analysis of the retrieved facts
  • Limitations — explicit coverage gaps or caveats before any total
  • Governed sources — the named tools invoked for this answer
  • Quality score — a deterministic 0–100 score covering structured output, governed-source grounding, partial-coverage disclosure, Markdown table validity, follow-up perspective, and summary completeness
Server-owned action links connect evidence back to the relevant Flux page. Follow-up questions are offered. Per-answer performance details show model, tool, DuckDB/report, application, validation, and combined network/render timing.

​
Access and retention

Transcript retention is for administrator quality review only. Administrators can inspect retained prompts, validated summaries, feedback, response modes, slow-request counts, and stage bottlenecks from the Intelligence Workspace quality review panel — without accessing model reasoning.

​
Provider configuration

Three provider adapters are available, selected with FLUX_AI_PROVIDER and switchable at runtime under Administration → AI: Key configuration variables:

​
Enabling Ask Flux

1

Set the feature flag

2

Choose a provider and supply credentials

Set FLUX_AI_PROVIDER to deepseek, openrouter, or foundry. Supply the corresponding API key as an App Service Key Vault reference — never as a plain environment variable in the deployment pipeline.
3

Configure the model (optional)

Override the default fast and deep model names if needed for your provider or procurement agreement.
4

Verify in Administration

Open Administration → AI in Flux to confirm the provider connection and check the spending ceiling configuration.
Every provider API key must be stored as an App Service Key Vault reference. Passing keys through appSettings in a pipeline parameter would wipe the reference on every deployment. Apply non-secret settings additively via az webapp config appsettings set to preserve Key Vault references.