Write extraction guidance the AI can follow
The per-field instructions that turn wrong or blank extractions into right ones — where the AI Extraction Guidance box lives, what good guidance sounds like, and the format guide and value lists that back it up.
At the end of this guide, the fields Paige keeps getting wrong have instructions that fix them — written the way the AI actually uses them: context and location, not just wishes.
Where the guidance lives
Open Settings (the gear at the bottom of the left menu — administrators only) and pick the scan job.
Choose Document Types, open the document type, and select the field in the Fields list.
Expand Advanced Settings. The box is AI Extraction Guidance — as its tooltip says, this "goes beyond format and focuses on context and location."
The document type editor with a field selected and AI Extraction Guidance expanded under Advanced Settings
What good guidance sounds like
The placeholder gives the pattern: "This field is always located below the company logo" or "The value should never exceed $100." Three ingredients, in the field's own words:
- Where it is on the page — position, nearby labels, which box on a form
- What it looks like — and what it is not, naming the values it gets confused with
- Rules a human checker would apply — ranges, relationships to other fields
Here's the shape at full strength, from a professionally-written example for a W-2's employer name:
"It is typically located in Box 'c' near the top of the form and is labeled as 'Employer's name.' … Focus on the name explicitly associated with the entity issuing the W2, avoiding confusion with the employee's name or other fields."
Location, label, and the specific confusion to avoid — that's the recipe. Guidance like "extract the correct value" adds nothing; guidance like "the ship-to address in the upper right, never the billing address beside it" fixes real mistakes.
The two supporting tools in the same panel
- Custom Data Format Guide (appears when you pick the custom option under the field's Data Format): describe the expected shape in words — the tooltip is explicit that this is "not a rigid mask, but a guide for AI."
- Force Selection from Embedded List: upload a text, CSV, or XML file of valid values, and extracted values are matched against it — the strongest fix for fields with a known universe (your vendor list, your GL codes). Save the document type before uploading the list; Replace List and Remove List manage it afterward.
Iterate like it's training — because it is
Change one field's guidance.
Run a few documents that previously extracted wrong.
Check the results, sharpen the wording, repeat.
Success check: the previously-wrong documents now extract correctly and a couple of previously-right ones still do — guidance that fixes one case by breaking another needs its "what it is not" clause tightened.
When guidance isn't the problem
Blank or wrong values sometimes come from earlier in the pipeline. If every field on a document is wrong, the document was probably classified as the wrong type — that's a classification problem, fixed in Teach Paige to identify document types. If a value from a different document appears, pages were grouped into the wrong document — that's separation. Paige extracted a blank or wrong field walks the full diagnosis order before you over-tune guidance.