Office documents are pictures with structure. Graph-Me-Maybe reads both.
Excel sheets, PowerPoint slides, Word pages and PDFs are drawn by a browser engine. Every cell edge, shape and line of text is recorded, so a place on the page can always be found again — and highlighted when it answers a question.
The model looks at each page and decides what is a heading, what sits under it, which note belongs to which table. Page numbers, watermarks and repeated footers are dropped. Pages are read side by side, not in a long queue.
Units, tables and charts become nodes in a Neo4j graph on your computer. A document that names another by number or title is linked to it as a citation; documents that say similar things are linked as similar.
Vector and keyword search run together; the documents found bring the ones they link to, up to three steps away. The answer cites each document, and opening it shows the cell, slide or paragraph highlighted.
Text extraction turns a sheet into a run of values. Rendering keeps which number sits under which header and beside which note.
Title, chart and caption mean something together because of where they sit on the slide.
A pasted screenshot, a floor plan, a diagram: there is no text to extract, but there is something to read.