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

# Complete document inventories

> Extract every requested item from a long text file while retaining fields and source evidence.

When you need every exercise, action item, or finding in a document, ask GAIA for
an exhaustive inventory and name the fields to retain:

```text theme={null}
Use document-extract to list every exercise in workshop.txt
with fields: name, reps, cue. Save the complete inventory to exercises.json.
```

Attach the source file in the Agent UI or provide a path permitted in your
session. The extractor accepts UTF-8 text files; convert other document formats
to text first. For an ordinary summary, ask for a summary instead.

## How it works

The default skill routes the task to `document-extract`. GAIA reads small,
overlapping sections sequentially and checks each section again for omissions.
Each retained item includes the requested fields, an exact source quotation,
and character positions. A missing field is marked `not stated`.

The framework appends the retained inventory to GAIA's answer, so the answer
can add analysis but can never condense forty extracted items into fifteen. For a requested file,
the exporter writes that same inventory as JSON, CSV, or readable text and checks
the saved contents before reporting completion.

Full page coverage does not guarantee that a model recognized every implicit
item. Review ambiguous passages and compare against a known item count when one
is available. When two extractions might be the same item, both are kept and the
later one is marked "(may repeat item N)" for you to check. Failed pages, denied
access, changed sources, and exceeded limits are reported as incomplete.

![Flagship Agent UI showing the complete two-item validation inventory with source references](https://mintcdn.com/amd-fe836e11/dmPGg26BI-9rXWdm/assets/document-extraction-validation.png?fit=max\&auto=format\&n=dmPGg26BI-9rXWdm\&q=85\&s=8d5e5e2b31ebca185e1cee09b868cac1)

## Memory and follow-ups

Outside private mode, existing memory records the extraction action, source
fingerprint, coverage, count, and final answer. Long extraction answers are kept
as one bounded conversation record so later items remain searchable. Ask GAIA
to search past conversations for details from a previous inventory.

A remembered inventory describes the earlier source. GAIA must revalidate the
current file before using it as evidence that a new extraction task is complete.
Private mode suppresses memory persistence.

## Limits and validation

Each source is limited to 256,000 characters, 512 occurrences, 100,000 characters
of retained item text and quotations, and 30 minutes of extraction time. Saved
conversation records are limited to 256,000 characters; a longer answer is not
saved to memory, and GAIA logs an error. Other limits end the extraction as
incomplete.
The backend must support JSON-schema responses; the Claude provider ignores the schema, so there each page relies on the prompt alone. Small local models can take
around 17 minutes for a 40-item, 79K-character document. Two passes per page
require at least 40 calls at that size; retries and slower models can exceed
the 30-minute deadline.

Developers can run the repeatable live benchmark with either a local model or a
configured Fireworks model:

```bash theme={null}
python scripts/dev/validate_document_extraction.py MODEL /tmp/extraction-validation
python scripts/dev/validate_document_extraction.py MODEL /tmp/extraction-unlined --unlined
```

The output directory must be new. This uses a synthetic transcript with forty
known items and checks every name, repetition count, and cue, the saved file,
and a later answer retrieved through the memory search tool. The `--unlined`
variant has no line breaks and starts with an entire page containing no items,
matching the format of transcription output.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.