Enterprise Document Intelligence Vol.1 #14C addresses structured extraction for collections with one document type repeated across many files — like thousands of insurance policies or purchase orders. Instead of letting models discover structure, the article advocates interviewing the person who files the documents to identify lookup fields (client, effective date, product, premium, renewal date, policy number). These fields were designed into the original form, making the collection an untyped database whose schema lives in the filer's head.

The piece covers the column interview process, two signals distinguishing real columns from fields that break filters later, the index as a growing table of contents, cost arithmetic (one-time per document vs. per-question costs), and the break-even point at fifty questions a day. It assumes the diagnostic from Article 14A confirmed this corpus shape.

A companion notebook in the public repo doc-intel/notebooks-vol1 lets users declare columns, fill them across look-alike PDFs, and measure fill rates — revealing fields that sound like columns but return half-empty. The worked example uses a fictional insurance portfolio; Section 3.2 provides a public folder for testing. This is Part IV of the series, where input shifts from file to folder and a table becomes the right answer.