Finding the sections. Each section starts where a title line such as SECTION 03 30 00 - CAST-IN-PLACE CONCRETE appears at the top of a page, and ends at END OF SECTION or the next title line. A table of contents page does not start a section, and a section title repeated as a running header on every page does not split one.
Clearing the page furniture. Project running headers, edition stamps, and page footers are set aside before the text is read, so a paragraph that crosses a page break reads as one paragraph. Footers are not thrown away: lines such as 03 30 00 - 3 or Page 2 of 12 become the printed page on each row.
Rebuilding the outline. Part headings such as PART 1 - GENERAL or PART ONE, articles such as 1.3 SUBMITTALS, and outline labels A., 1., a., and 1) are rebuilt into a path, so a row is labeled 1.3.C.1 rather than just 1. UFGS decimal paragraphs such as 1.3.1 are read the same way.
Deciding what is a requirement. A paragraph becomes a row when it opens with Submit, when it starts with a type label such as Shop Drawings: or Product Data:, or when it names a submittal type inside a Submittals article. A paragraph elsewhere that uses the verb submit, or asks to provide, furnish, or include a named type, becomes a row with a lower confidence hint. The word submittals in running prose, such as in accordance with approved submittals, does not.
Naming the type. Keywords match whole words: shop, coordination, fabrication, and placing drawings are Shop Drawings; product data, manufacturer’s literature, data sheets, and catalog cuts are Product Data; samples and mock-ups are Samples; calculations and delegated design are Design Data; test reports and mix designs are Test Reports; certificates and qualifications are Certificates; warranties, record documents, spare parts, and extra stock are Closeout. The earliest keyword in a paragraph wins, and a label before a colon wins over the body.
Action or informational. The Action/Info column comes from the article title: Action Submittals give Action, Informational Submittals give Informational, and Closeout Submittals give Closeout.
Scanned pages. Pages with no text layer are read with the OCR language installed in Windows, then parsed like any other page.
What each row carries. Spec section, section title, part, article, paragraph, PDF page, printed page, the requirement text up to 400 characters, type, action, and a confidence hint from 0 to 1, ready to review beside the page and export.