Architecture

The app is a FastAPI backend with a React frontend, deployed as one Databricks App. It never imports the framework itself: it turns a report definition into an Impulse notebook and runs that notebook as a serverless job with a pinned Impulse release.

Components

Component	Role
Frontend (frontend/)	React and Vite: the report steps, the results charts and the chat window
Backend (app.py, server/)	FastAPI REST API; finds tables, units under test and channels through Unity Catalog and the SQL warehouse
Report definition	Data, units under test, signals, events and aggregations, checked against the framework rules as the user edits
Code generation	Turns the report definition into one Impulse notebook with the configuration, signals, events and aggregations
Runs	Pre-flight check of the user's and the service principal's access, then a serverless job that installs the pinned Impulse wheel
Results	Queries the report's gold tables with the user's permissions
Assistant	A Foundation Model API chat model with tools that do what the UI does, guided by app skills and the framework skills of the pinned version

Identities

• The signed-in user. Through Databricks Apps user authorization, every catalog, SQL, workspace and model call runs with the user's own permissions, so users only see data they can read.
• The app's service principal. Report jobs run as the service principal, which needs access to the silver and gold schemas. Before each run the app checks that both the user and the service principal can read the source tables and create tables in the output schema, and it shows the missing GRANT statements otherwise. The user gets CAN_MANAGE_RUN on their report's job.

How a report runs

1. The report definition is saved in the user's workspace folder.
2. The app generates the notebook and runs the pre-flight check.
3. The notebook is published to the service principal's workspace folder, readable by the user.
4. A serverless job per user and report runs it, installing the Impulse wheel pinned by the app.
5. The job recomputes the selected recordings and overwrites the report's gold tables.
6. The app reads the gold tables back for the results charts.

Where things are stored

What	Where
Reports and their chat history	/Users/<user>/Impulse/<report>/
Each user's settings (SQL warehouse, chat model)	/Users/<user>/Impulse/settings.json
The shared data model	/Workspace/Shared/Impulse/config.yaml
Generated notebooks	/Users/<service principal>/Impulse/<user>/<report>/report
Report jobs	Impulse <report> (<user>), tagged with the app, report, user and Impulse version
Results	Gold tables <catalog>.<schema>.<table prefix>_* in Unity Catalog (the prefix defaults to the report name), see Gold Layer Schema

Framework version

impulse.lock pins one Impulse release: the wheel that report jobs install, the serverless environment version, and the framework skills the assistant reads. The app's tests check every generated notebook against the same release.