AI extraction & instructions

AI extraction reads each file and returns one value for every extracted input you define, and only those. A field can define more than one, each with its own rule. When a value comes back wrong, add an instruction to steer the next run.

Updated 3 min read

The AI maps the text and visuals of a file onto your schema. It leaves unfound items empty and never invents data. Extracted values land in case data under stable keys. Renaming a field's display label later does not disturb your stored data.

How the app selects data

Extraction uses one call per template run. The app sends a field to the model when it has an extracted input and at least one signal. Signals include an AI instruction, a read zone, or an AI instruction output. When a field qualifies, every extracted input on it goes to the model. Bare fields are skipped. Form field and Static inputs are not extracted.

Reducing work means fewer extracted inputs or a separate template.

You choose the model in Settings > Preferences at the AI model row. Gemini 2.5 Flash Lite is the default and lowest cost. Gemini 3.1 Flash Lite has a higher cost. Your choice covers extraction, autofill, and document reflows. An admin can restrict which models appear. The chat drawer has its own selector.

The cloud ingestion path sends the PDF itself. The dashboard Re-extract path transcribes the file to text first. This path drops vision and OCR hints.

Value types and read zones

The Date value type forces the model to return a date formatted as YYYY-MM-DD. The app keeps the document's literal wording as metadata. The model ignores Select options and Number types. You control number formats through AI instructions or a Computed output.

You can add read zones to an extracted input. They carry two toggles. Extract value from selected region reads visual elements like signatures or photos. OCR (reads whole page) reads scanned pages.

An input's AI instruction replaces the printed label as the model's hint. The instruction must state what the value is. Changing an input's Value source clears its AI instructions. You can use undo to restore them.

The four instruction layers

Push every rule to the lowest applicable layer. A value's own rule always wins for that value. Template rules apply to every value and act as the default.

1Open the field's extracted input→2Add a specific AI instruction→3Re-extract one case→4Iterate one rule at a time
LayerWhere you edit itScope
ValueThe AI instructions box on an extracted input in the Inputs section.One value, every extraction.
TemplateThe AI instructions section behind the Data tab's chevron.All values. It appears only when the template has an extracted input.
Per-sourceThe AI instructions textarea in a document tab's chevron popover.That document in this template. It guides extraction and the page-shift interpreter.
SpecialThe review modal's Special AI Instructions box, the Run button's caret menu, or Add AI instructions.One run only.

Special instructions override template rules for any value they name. The app discards them when a modal closes. The template editor saves them in the Test & Preview cache.

You find special instructions in the review modal Special AI Instructions box. In the template editor they live in the Run button's caret menu under AI instructions for this run. The caret shows an accent dot when text is entered. You can also use Add AI instructions in the Documents console. In the console, the text only takes effect when the run performs an extraction.

A packet using three layers
ValueReference number: the code after REF, not the page number
Per-sourcePage 2 can spill onto an extra sheet
TemplateAmounts are in USD

Advanced behaviors

For a Batch field, name the column position in your instructions.

Sample learning builds a document profile. The app sends this profile to the extraction call alongside your rules.

AI instruction outputs run after extraction completes. They derive new data from the already-extracted values. Your template rules thread into these derivations.

When you edit a value, its tag flips to User Input. Re-extract skips these user-written keys. If a value is too difficult for extraction, change its Value source to Form field.

A file becomes extractable when a template runs on it. Photos and receipts stay untagged and never become extraction targets.

Type at least 2 characters to search.
We use cookies to keep you signed in and improve the app. See our Cookie Policy.
Manage preferences