Zava Service Desk — Fabric IQ × Foundry demo

One trusted context for an AI-run service desk: Fabric computes how far the XLA fell, the contract says what Zava owes.

Synthetic data. Zava is a fictional managed-service operator. Its customers, users, devices, tickets, contracts and clauses are generated from seed 42 by fabric/data/generate_data.py out of fabric/data/world.yaml. No real customer, GUID or endpoint appears anywhere in this repository.

AI agents resolve tickets on their own. Running a whole service desk on them — six customers, under contract — needs one trusted context. This demo answers a question that neither the data nor the contract can answer alone:

"Fabrikam and Litware both fell below the same 40% zero-touch XLA last week — what does each contract make Zava owe?"

📊 The figure comes from Fabric: DAX over a Direct Lake semantic model, the ontology and the Eventhouse, queried by a Fabric Data Agent.
📜 The obligation comes from the contract: a Foundry contracts agent retrieves and cites the clause from a Foundry IQ knowledge base.
🧭 The context comes from Work IQ and Web IQ: who already acts, what was announced.
🚫 The language model never computes. It routes, cites and phrases.

At a glance


Domain	Managed IT service desk — tickets, digital employee experience (DEM), AI agents, XLAs
Company	Zava (fictional operator), six fictional customers: Contoso, Fabrikam, Litware, Northwind, Tailspin, Woodgrove
Fabric items	Lakehouse · ingestion notebook · Eventhouse · Ontology + graph (Fabric IQ) · Semantic model · Report · RTI dashboard · Activator · Operations Agent · Data Agent (MCP) · Task flow · Rayfin app
Foundry plane	Supervisor agent · contracts agent · Foundry IQ knowledge base · Work IQ · Web IQ
Hosted console	Rayfin Fabric App · React + Vite · single-tenant Entra SPA, delegated scopes, no secret
Runs without a tenant?	Data generation, live-stream replay, tests and the console preview: yes. Deployment: no.

What is live and what is staged

Plane	State	Where it runs
Fabric — Lakehouse, Eventhouse, ontology, semantic model, report, RTI, Data Agent	🟢 Live	Deployed by `deploy_all.py` into your workspace
Console — portfolio, customer pages, assistant	🟢 Live	Evaluates named DAX measures; the assistant calls the Data Agent
Foundry — supervisor, contracts agent, Foundry IQ, Work IQ, Web IQ	🎬 Staged	Played in the console from `scenario.json`, no Foundry resource

The screen carries no label about what is staged — the presenter owns that framing, as written in the demo script.

The storyline

A Corporate VPN client update breaks remote access at Fabrikam Lyon. Tickets pile up, the AI agents escalate instead of resolving, and the zero-touch rate falls through the XLA. In the same week Litware misses the same target for a different reason. Their numbers are close; their contractual consequences are opposite.

Customer, ISO week 38	Zero-touch (previous week)	XLA	Contract	What Zava owes
Fabrikam Industries	34.0% (46.0%)	≥ 40%	CTR-FAB-2026 · XLA-FAB-01	A 9,250 EUR service credit — 5% of the 185,000 EUR monthly fee, capped at 15%
Litware Insurance	38.0% (48.0%)	≥ 40%	CTR-LIT-2026 · XLA-LIT-01	No credit. A written remediation plan within 10 business days — due 2 October

These figures are asserted by the test suite and reproduced by generate_data (see Quick start). At deploy time --shift-weeks auto moves the history so the breach always lands in the last closed week.

An XLA (eXperience Level Agreement) commits to what users live through — zero-touch rate, CSAT, device experience, time lost — where an SLA only commits to technical targets.

The boundary rule

Layer	Owns	Never does
Fabric (semantic model, ontology, Eventhouse, Data Agent)	Every measure, aggregation and threshold	—
Foundry (supervisor, contracts agent, Foundry IQ)	Orchestration, clause retrieval and citation, the written answer	Reimplement a metric
Work IQ · Web IQ	Who already acts, what was announced publicly	Produce a figure
Console (TypeScript)	Evaluates named DAX measures, renders them	Compute a rate, a credit or a threshold

The Fabric hop costs latency — the Data Agent takes 40–160 s per answer. That cost is accepted on purpose: two definitions of "zero-touch", one in Fabric and one in a prompt, would be unauditable.

How it fits together

flowchart LR
  subgraph src[Synthetic sources]
    CSV[ITSM, DEM, CSAT, contracts<br/>reference + 90 days]
    LIVE[Live streams<br/>telemetry, traces, tickets]
  end

  subgraph fabric[Microsoft Fabric]
    NB[Ingestion notebook]
    LH[(Lakehouse<br/>LH_ServiceDesk)]
    EH[(Eventhouse<br/>KQL_ServiceDesk)]
    ONT[Ontology + graph<br/>ONT_ServiceDesk]
    SM[Semantic model<br/>SM_ServiceDesk_Analytics]
    DA[Data Agent<br/>ServiceDesk_Analyst]
    RTI[RTI dashboard · Activator<br/>Operations Agent]
  end

  subgraph foundry[Microsoft Foundry — staged in this demo]
    SUP[Supervisor<br/>Zava-ServiceDesk-Agent]
    CTR[Contracts agent<br/>Zava-SD-Contracts]
    KB[Foundry IQ<br/>Service agreements]
    WIQ[Work IQ<br/>mail, Teams, meetings]
    WEB[Web IQ<br/>public announcements]
  end

  CSV --> NB --> LH
  LIVE --> EH
  LH --> ONT
  LH --> SM
  EH --> ONT
  EH --> RTI
  ONT --> DA
  SM --> DA
  EH -->|KQL| DA
  DA -->|MCP tool| SUP
  KB -->|retrieval| CTR
  CTR -->|A2A| SUP
  WIQ -->|MCP| SUP
  WEB -->|tool| SUP
  SUP --> APP[Rayfin Fabric App<br/>Zava Service Desk console]
  SM -->|DAX| APP

  classDef vision stroke-dasharray: 5 4
  class SUP,CTR,KB,WIQ,WEB vision

Dashed nodes are staged. Two attachment kinds, not interchangeable: the Data Agent is a tool — the supervisor delegates the question and Fabric returns an answer. The contract corpus is a knowledge source — text comes back and the contracts agent reasons over it. The full rationale is in docs/ARCHITECTURE.md.

Screens

Screen	What to look at
	Zava IQ. One prompt per IQ layer — Fabric IQ, Foundry IQ, Work IQ, Web IQ — each answer citing its source, then a recommendation and the Finance approval. The page scrolls with each click.
	Portfolio. Every figure is a named DAX measure evaluated live against the semantic model — Fabrikam and Litware show red against their 40% target.
	Architecture. The chain from synthetic sources to the console, with the boundary between Fabric and Foundry.

Repository structure

Deployment code is grouped one package per Fabric workload — the folder tells you which artifact the code produces.

Path	What lives there
`deploy_all.py`	One-shot idempotent orchestrator, plus the pre-demo warm-up
`fabric/`	One package per workload: `_shared`, `data`, `workspace`, `lakehouse`, `eventhouse`, `ontology`, `graph`, `powerbi`, `rti`, `data_agent`, `taskflow`, `app`
`app-zava-service-desk/`	The Rayfin console — React + Vite, live DAX, Data Agent assistant, Zava IQ
`docs/`	Architecture, deployment runbook, engineering notes, demo script, screenshots
`tests/`	The offline gate, run before every deploy
`scripts/`	Repository leak guard, also run in CI
`config.example.yaml` · `state.example.json`	Templates for the git-ignored tenant configuration

Regenerate the full file list with git ls-files — it is not duplicated here, so it cannot drift.

Prerequisites

Need	Version	For
Python	3.12	Data generation, tests, deployment — packages in `requirements.txt`
Node.js	20+	The console (`npm ci` in `app-zava-service-desk/`)
Microsoft Fabric capacity	F-SKU or trial, running	Deployment only — a paused capacity answers 404
Azure CLI	signed in with `az login`	Deployment only — the deploy refuses any account other than `deployment.expected_account`
Entra permission	create an SPA app registration	Deployment of the console only

Quick start

Offline first — no tenant, no network.

git clone https://github.com/Statyx/Fab-ServiceDesk-IQ.git
cd Fab-ServiceDesk-IQ
pip install -r requirements.txt
python -m fabric.data.generate_data        # writes artifacts/data/, byte-identical everywhere
python -m pytest tests -q                  # the offline gate

generate_data ends with the storyline, which is how you know the dataset is right:

XLA breaches (closed windows):
  CUS-FAB XLA-FAB-01 zero_touch_rate ...: 34.0 vs 40.0 → credit 9,250.00 EUR
  CUS-LIT XLA-LIT-01 zero_touch_rate ...: 38.0 vs 40.0 → credit 0.00 EUR

Replay the live incident offline (writes the streams to files instead of the Eventhouse):

python -m fabric.eventhouse.inject_event --scenario vpn-lyon --loop --cycles 6 --out artifacts/live

Run the console from fixtures:

cd app-zava-service-desk
npm ci
npm test
npx vite                                   # then open /preview — every screen, no tenant

Deploy to Fabric

Every step is idempotent: it finds its item by name and creates only what is missing.

cp config.example.yaml config.yaml         # fill the capacity and deployment.expected_account
python deploy_all.py --list                # the steps, in order
python deploy_all.py                       # every step, then a warm-up

Option	Effect
`python deploy_all.py ontology graph`	Run only the named steps
`python deploy_all.py --from ontology`	Resume after a failure
`python deploy_all.py --skip app,taskflow`	Run everything except these
`python deploy_all.py --no-warmup`	Skip the final warm-up
`python deploy_all.py --warmup`	Warm-up only — run it right before the demo to pay the cold start off-stage
`python -m fabric.app.deploy_app`	Redeploy the console alone

Deploy only what changed. A console change needs deploy_app, not a full run.

The steps, in order: generate_data → workspace → lakehouse → setup_notebook → eventhouse → preload → ontology → graph → semantic_model → verify_model → report → dashboard → activator → operations_agent → data_agent → taskflow → app.

Configuration can also live in one folder per tenant under deployments/<name>/ — all of it git-ignored. Three steps stay UI-only: start the Activator, bind the Operations Agent, import the task flow. The deployment runbook lists them, with the MCP calls and the recorded answers.

Quality gate

A green suite is the definition of done.

python -m pytest tests -q
python scripts/check_repo_leaks.py         # stage new files first — it scans tracked files
cd app-zava-service-desk && npx vitest run && npm run lint && npm run build

CI (no-client-leak.yml) runs the same gate on every push — leak guard, dataset smoke run, pytest, console lint/test/build — and refuses tracked generated data, local configuration and Windows-only scripts.

Troubleshooting

Known failure modes — Fabric REST quirks, ontology and Data Agent traps, RTI, Rayfin — and the workaround that holds each one are in docs/ENGINEERING-NOTES.md. The most common:

Symptom	Fix
404 on every Fabric call	The capacity is paused — resume it
The Data Agent is slow on the first question	Run `python deploy_all.py --warmup` before the demo
The deploy stops on "Signed in as …, profile expects …"	`az login` as `deployment.expected_account` from your config
A step failed half-way	Fix the cause, then `python deploy_all.py --from <step>` — finished items are reused

Contributing

The test suite enforces most of these rules; read .github/copilot-instructions.md for the full list.

Layout. A new deploy script goes in the folder of the artifact it creates, as deploy_<artifact>.py, and is registered in deploy_all.py. Shared plumbing goes in fabric/_shared/ only if it knows about no specific artifact.
Run as modules from the repository root: python -m fabric.ontology.deploy_ontology.
Prologue. Every runnable module calls fabric._shared.platform_env.bootstrap() before any third-party import. winreg is imported in platform_env.py only.
No secrets, no IDs. No shell=True, no hard-coded GUID — identifiers live in the git-ignored config.yaml / state.json or deployments/<name>/. rayfin/rayfin.yml stays tenant-neutral.
Idempotent. Find by name, create only what is missing.
Deterministic data. Seed 42 from world.yaml; the storyline figures are tested.
Boundary. Never compute a rate, a credit or a threshold in a prompt or in TypeScript.
UI wording. Screens never label the storyline as simulated or offer a source-mode selector; the framing belongs in the demo script.

Open a pull request once the quality gate is green.

Documentation

Document	For
`docs/ARCHITECTURE.md`	Why the pieces are wired this way: data model, ontology, storyboard
`docs/DEPLOYMENT.md`	Deploy steps, items, UI-only steps, the console, `state.json`
`docs/ENGINEERING-NOTES.md`	Failure modes and the workarounds that hold them
`docs/demo/DEMO_SCRIPT.html`	Screen by screen, what to say, live vs staged, plan B

License

MIT. See LICENSE.

Navigation Menu

Repository navigation

About

Repository files navigation

Zava Service Desk — Fabric IQ × Foundry demo

Contents

At a glance

The storyline

The boundary rule

How it fits together

Screens

Repository structure

Prerequisites

Quick start

Deploy to Fabric

Quality gate

Troubleshooting

Contributing

Documentation

License

About

Releases

Packages

Contributors

Languages

Name		Name	Last commit message	Last commit date
.github		.github
app-zava-service-desk		app-zava-service-desk
docs		docs
fabric		fabric
scripts		scripts
tests		tests
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
config.example.yaml		config.example.yaml
deploy_all.py		deploy_all.py
requirements.txt		requirements.txt
state.example.json		state.example.json

Latest commit

History

Folders and files

Repository files navigation

Zava Service Desk — Fabric IQ × Foundry demo

Contents

At a glance

The storyline

The boundary rule

How it fits together

Screens

Repository structure

Prerequisites

Quick start

Deploy to Fabric

Quality gate

Troubleshooting

Contributing

Documentation

License

About

Topics

Resources

Contributing

Security policy

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages