An extracted label is a model output under a contract. It is not an independently verified fact.

Extraction contract listing fields and coverage
Tags you can trust are approved contract fields. Coverage is field health, not a fact check.

Metadata and tags

Classify one fictional note into a closed set of labels, then read the field back. Hands-on time is about 25 minutes after the file is ready. Extraction waits are extra.

Purpose

You can tell a file name from an extracted field. Level: editor who can approve a schema. Role: schema edit plus extraction run. Mode: stored. Start from Upload files.

Prerequisites

The note harbor-note.txt is on the collection and has reached ready. An organization generation key exists only if the extraction model needs one. Viewers cannot approve the schema.

Sample

Schema field delivery_window_days, type number, optional. Enum field review_label with values standard, rush, unknown. The note says 14 days, so a careful extraction may set standard or leave unknown when the text does not use those words. Expected empty behavior: null or omitted means the model did not assert a value. It does not mean the value is zero.

Steps

  1. In Schema Studio, add those two fields and approve the schema version you will run. A draft schema does not extract.
  2. Run extraction on the sample scope that includes harbor-note.txt. Record the run id.
  3. Open Contract on the collection. Read review_label and delivery_window_days for that document.
  4. If the label is outside the enum, treat the run as failed validation. Do not hand-edit a sidecar file. There is no metadata write API.

Fields

NameLevelMeaning
File name, format, relativePathDocument rowProperties from upload. Not extracted tags.
statusDocument rowprocessing, ready, extracted, indexed, error, and the states between them.
Contract fieldApproved schemaThe only tag a later search should filter on.
Null or unknownField valueNo assertion. Do not coerce null to a default label.

Example

A review record you keep outside the product, not a Bundata request:

{"document":"harbor-note.txt","review_label":"unknown","delivery_window_days":null,"reason":"The note states 14 days and does not use the words standard or rush."}

Recovery

If extraction returns error, open the run and retry that run. Do not upload the file again. If the queue code is DOCUMENT_PROCESSING_QUEUE_UNAVAILABLE, retry the same confirmation from the upload recipe. A confident wrong label is still not a fact: send it to a person.

Limits

No .meta.json ingest, no arbitrary metadata editor, and no claim that a tag was verified by a second system. Folder paths are not tags.

Check

Success: Contract shows the field with a value inside the enum or an explicit empty. Failure: a label that is not in the enum is rejected by your review, and you do not store it as approved.

Related documentation