Every citation in InsightDealDesk — on a rep, a finding, a copilot answer — points at a piece of a document that knows its page. Getting documents into that shape is the reader’s job.
Two readers, one contract
In production, documents uploaded to the data room go to object storage and are read by InsightXtract — OCR, layout and tables — which calls back with fields and citations. On a laptop, in tests and on the illustrative book, a built-in reader produces the same shape from the document’s text. Both return page-level pieces: an SPA section with its number and page, a schedule item with its schedule reference, a diligence finding with its severity and the reps it covers.
configured?"} X -->|yes| XT["InsightXtract: OCR,
layout, tables"] X -->|no| BI["Built-in reader"] XT --> PC["Pieces: kind, ref,
heading, page, text"] BI --> PC PC --> RP["Reps table"] PC --> FD["Findings"] PC --> SR["Search and copilot"]
Uploads never touch local disk
Without an object store configured, uploads are refused rather than written to the server’s disk. A data room is client-confidential; it lives where the brokerage keeps confidential files.

