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

# AI document classification

> Let AI classify an uploaded document and fill its keywords in one step, with a confidence score and the reasoning behind every suggestion.

<Note>
  AI document classification is under active development. Suggestion quality depends on how the
  classification tree is configured for your environment, and features or UI labels can change
  between releases.
</Note>

## Overview

AI document classification reads an uploaded document and suggests both its **document type** and its
**keyword values** in a single step. On the upload page it appears as one action:

* **Per file** — the robot button on a file row analyzes that document.
* **For all files** — the **AI Index** button in the toolbar analyzes every eligible file in
  parallel.

There is no separate "guess the type" or "guess the keywords" action: one analysis produces both,
from one reading of the document.

## How a suggestion is produced

Every suggestion comes from the [classification tree](/document-classification/how-it-runs) your
administrator authored. The AI walks the tree — a series of small, controlled decisions — until
it lands on a document type, then extracts that type's keywords using the instructions collected
along the way. The tree's branches define the complete set of types the AI may suggest.

**The tree is the feature.** Without a configured tree the analysis makes no suggestions at all —
there is no generic fallback guess. Author the tree first; see
[Configuring the tree](/document-classification/configuring-the-tree).

The suggested type is applied to the file row, the keyword grid is rebuilt for that type, and the
extracted values are filled in. Values you have already entered by hand are kept.

## Confidence and reasoning

Every suggestion carries a **confidence score** and the **reasoning** behind it. Hover the
percentage next to the file row to see:

* The classification path the document took through the tree, step by step.
* The evidence the AI quoted from the document at each step.
* The average confidence of the extracted keywords.

The score is deliberately conservative: it equals the *weakest* step of the classification, and
it is capped at 50% when the walk could not settle on a specific sub-category. A step whose
quoted evidence could not be found in the document is flagged in the tooltip. A low score or a
flagged step is a signal to double-check the suggestion before uploading.

## When you get no suggestion

The analysis returns nothing — no type, no keywords — rather than a doubtful guess when:

* **No classification tree is configured.** The tree defines everything the AI is allowed to
  suggest; until it has branches, the analysis has nothing to suggest from.
* The document matches **no branch** of the classification tree. A document outside the tree is
  left for you to index manually.
* The suggested type is one **you are excluded from** (your administrator has hidden certain
  document types from your upload page). The AI never suggests a type you cannot pick.
* Two classification attempts **disagreed** about where the document belongs — see
  [self-checking](/document-classification/how-it-runs#halted-walks-and-the-automatic-re-walk).
* No text could be read from the document at all.

In these cases pick the document type yourself and fill the keywords manually.