Document Extractor

Pull the fields you NAME out of a document — an invoice, a transcript, a clinical note, a policy, a contract — and get, per field, whether it was found, the character offset and line it came from, and whether more than one answer was available. A field is found, ambiguous or not_found, and NEVER an empty string. That is the whole point. Measured across the ten extraction agents this consolidates: nine of them have zero references to a not-found or unextracted state, while carrying between five and twenty-one empty-string defaults. So a field the tool could not extract came back indistinguishable from a field that was present in the document and genuinely blank. On an invoice that is a total nobody could find, reported as no total. On a clinical note it is a measurement nobody took, reported as a measurement. AMBIGUOUS IS THE STATE THEY HAD NO ROOM FOR AT ALL. When your labels match more than once with different values, every candidate comes back and none is chosen. An invoice with a "Total" in the line-item block and a "Total" in the summary has two answers, and picking the first is a decision you did not authorise. The same value found twice is one answer stated twice, and is not ambiguous. EVERY FOUND VALUE CAN BE CHECKED. Each carries the character offset and the line it was taken from, so you can verify it against the document rather than trust it. It also reports which label actually matched, and flags a match on a label other than the field's own name — a field found through a label you did not expect may not be the field you meant. Two refusals worth knowing. A not_found is a statement about the labels you supplied and the text you supplied, never a claim that the document lacks the information. And labels are never defaulted to the field name: searching for a label you never said appears in the document would report a not_found about a guess. It reads only the text you paste. Nothing is fetched, no field you did not name is inferred, and no value is normalised into a type you did not ask for.