The input is a document set, not one file
The customer email may define delivery, a PDF may contain specifications and an Excel sheet may contain quantities. Extraction must treat these sources as one case while retaining their individual provenance.
Extract fields with evidence
Useful output includes item descriptions, quantities, units, standards, materials, ratings, approved makes and due dates. Each value should carry a reference to its source. Low confidence or conflicting fields should enter a review queue.
Normalize carefully
Normalization makes supplier matching possible, but it can also hide meaning. Preserve original text beside normalized units and terminology. A reviewer should always be able to see what the customer actually wrote.
Measure corrections, not just extraction rate
A pilot should track missing requirements, wrong values, unsupported assumptions and reviewer edits. Evidence coverage and corrections per line provide a more honest picture than a single accuracy percentage.