Docs menu
AI assistants
What each assistant does, what it relies on, and what it never does on its own.
In short
Northstar has several AI assistants: Chat, the ingestion and transformation agents, assistants for alerts, SQL, charts and dashboards, and AI texts on quality, KPIs and the Smart Cockpit. They work on your governed datasets, within your rights, and they propose: you decide what is deployed, applied or accepted. Platform administrators choose which AI features are on.
Principles
- Grounded in governed data. Assistants work from your datasets, their documentation and their quality analysis. Chat shows how each answer was built: datasets, SQL and documents.
- Within your rights. An assistant only reads what you can read. Chat only uses datasets where you have the Ask AI right; the queries it writes, and those of dashboards and the SQL Executor, are checked table by table.
- They propose, you decide. The ingestion agent never deploys a pipeline. The transformation agent never applies a dataset. Suggested alerts, validation rules, KPIs and column descriptions wait for your review.
- Labelled. AI text is marked as such, for example AI suggestion on a column description, until someone accepts it.
- Generated code stays in its lane. The transformation agent only writes queries that read data; the statements that create views and tables are generated by Northstar.
The assistants
| Where | What it does | What it relies on | What it never does alone |
|---|---|---|---|
| Chat | Answers questions in plain language | Datasets and collections you can use with AI, gold first | Change data, or read a dataset you cannot use |
| Sources: ingestion agent | Writes and tests a pipeline that loads a source | Your description and the source's documentation | Deploy, or see the credentials you enter |
| Sources: proposals | Proposes silver and gold datasets after an ingestion | The source's bronze tables | Build anything: you open a session from the proposals |
| Datasets: transformation agent | Writes the queries of silver and gold datasets | Schema and sample rows of the input datasets | Apply, or write the statements that create tables |
| Dashboards: dashboard assistant | Adds, changes and resizes tiles | Datasets you can use; every query is checked | Show a tile to a reader without access to its data |
| Datasets: alert assistant | Prepares an alert from one sentence | The dataset's columns, validations and quality score | Create the alert without your review |
| KPIs, alerts, SQL Executor: SQL assistants | Write a query from a question, fix a failing query | The dataset's schema | Run the query without you |
| Builder: chart assistant | Suggests a figure from a sentence | The dataset's schema | Save the figure |
| Datasets: column descriptions | Describes columns | Documentation, column lineage and sample rows | Replace a description: it stays a suggestion until accepted |
| Datasets and KPIs: recommendations | Suggests validation rules and KPIs | The dataset's documentation and existing KPIs | Add a rule or a KPI without your review |
| Quality: AI summary | Explains each analysis | Aggregated statistics only | Read raw rows |
| Smart Cockpit | Writes KPI insights and the executive summary | Your KPIs' values and trends | — |
| Dashboards: AI summary | Summarises what a dashboard shows | The snapshot of the data displayed | — |
AI and dataset layers
Some AI features only run on curated datasets:
| AI feature | Layers |
|---|---|
| Validation rule recommendations, AI quality summary | Silver and gold |
| KPI recommendations | Gold |
| Chat | Gold first, then silver, then every dataset you can use |
See AI features by layer.
Turning AI features on and off
Platform administrators turn each AI feature on or off in Platform settings → AI features. A change applies to every organization within a few seconds.
The switches are: Smart Cockpit, Chat, Data quality summary, Dashboard summary, Dashboard assistant, SQL Executor assistant, Alert and KPI SQL assistants, Alert assistant, Builder assistant, Metadata and recommendations, Ingestion agent and Transformation agent.
When a feature is off, it is no longer offered; the rest of Northstar works as before.
What is sent to the AI
AI features call a language model configured for your platform. What each one sends:
| Feature | Sent to the model |
|---|---|
| Chat | Your question, the schemas of the datasets it uses, the results of its queries, and document excerpts |
| Quality summary | Aggregated statistics only, never raw rows |
| Column descriptions | Documentation, column lineage and a sample of rows |
| Transformation agent | Schemas and a sample of rows of the input datasets |
| Ingestion agent | Your messages and the documentation it reads. Never the credentials. |
| SQL Executor summary | The first rows of the page shown, only when you ask |
| Dashboard summary | The snapshot of the data displayed |
| Smart Cockpit | KPI values and trends |
If a dataset holds data that must not leave your warehouse, ask your administrators which AI features to keep on.
FAQ
Can an assistant see data I cannot see? No. Assistants act with your rights. A platform administrator's assistant sees what a platform administrator sees.
Why is an AI button missing? The feature may be off for the platform, the dataset's layer may not allow it, or your role or team may lack the right.
Is an AI answer always right? No. Answers are grounded and traceable, so you can check them: open How this answer was built in Chat, or the dataset behind a KPI.