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

​
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:
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.