Fable
The archive room first breaks the paper scrolls into entries that can be recorded
The county office storeroom was piled with contracts, tax receipts, legal complaints and field boundary maps. If the clerk compressed each volume into a string of text, titles, table columns, page-seam seals and small marginal notes would all be lost, and later retrieval could not return to the original page.
The clerk first determines the page structure, distinguishing headings, body text, annotations, tables, seals and captions, and preserves the page number, coordinates and original file source for each fragment. Scanned documents first go through OCR, then low-confidence regions are marked.
The row-column relationships in a tax invoice determine the meaning of the amounts. The clerk recorded cells, cross-row headers, and total relationships separately, avoiding lumping 'tax rate' and 'tax amount' into one vague block of text.
The archivist used sampled originals to verify text, structure, and source, and when a stamp was missed in recognition, only the corresponding page was corrected. Parsing failures enter a pending queue rather than silently generating a seemingly complete document.
The output of document parsing is not plain text, but indexable elements with structure, layout and provenance; this information directly determines whether RAG can find and cite the correct evidence.