> ## Documentation Index
> Fetch the complete documentation index at: https://docs.a7manager.se/llms.txt
> Use this file to discover all available pages before exploring further.

# Ask AI

> Natural-language questions answered against the platform's own data.

Ask AI answers questions about a project's data in plain language — in English or Swedish —
and returns prose, tables and charts. It is an exploration tool for questions that would
otherwise mean building a report.

## Principle

The question goes to a language model that has been given a set of **read-only query tools**
over the platform's data. The model decides which tools to call, the platform runs them
scoped to the selected project, and the model writes up the answer.

It does **not** write SQL and it cannot modify anything. The tools accept only an allowlisted
set of fields to group, filter and measure by, and the query layer refuses anything that is
not a read.

## Tools available

Nine tools are available: aggregation, trends over time, comparison between entities, detail
lookup, listing values, duration analysis, distributions, workflow analysis and maintenance
compliance.

They can be pointed at defects, warranty claims, maintenance records and tasks, vehicles,
components, vehicle configurations, out-of-service periods, cleaning records, measurements,
inspection reports, inspection findings and vehicle readings.

<Note>
  **Coverage lags feature development.** Modules and fields added recently may not yet be
  reachable through Ask AI, and the tool set is extended as the platform grows. If a question
  about a newer area comes back thin, that is the likely reason rather than a data problem.
</Note>

## Response shape

Answers come as a mix of prose, data tables and charts. Charts are produced by the model
itself as chart specifications, which the interface renders — the model is instructed to
chart anything with a time dimension and never to truncate the data behind a chart.

Conversations are **multi-turn**: follow-up questions keep the earlier context, so you can
refine rather than restate.

## Limits

| Limit                   | Value                                                             |
| ----------------------- | ----------------------------------------------------------------- |
| Conversation length     | 40 messages, after which a new conversation is needed             |
| Requests                | 100 per user per hour                                             |
| Message length          | 4 000 characters                                                  |
| Tool calls per question | 10 rounds, after which the model is made to summarise what it has |

A question is answered in a single request with a timeout under a minute, so very broad
questions are better split up.

## Scope and privacy

Every query is **hard-scoped to the one selected project**. There is no cross-project
querying, and the model cannot reach data outside the project you are working in.

Conversations are held in the browser for the duration of the session and are not stored
server-side. There are no saved or bookmarked answers.

## Model

The platform runs on Azure OpenAI, currently a GPT-5.4-mini deployment, configured per
environment rather than hard-coded.

An upgrade to a newer model — GPT-5.6, Terra or Luna — is planned. Because the deployment is
a configuration value, that change does not require an application release.

## Roles

Ask AI requires the **AI Analytics User** role on the project, which is granted separately
from the other roles precisely because it is a broad read across the whole project's data.
Project Admins have it by default.