Extracted text of a workspace PDF, page by page.

This is what the PDF text view renders, and it is deliberately the same pypdf extraction that knowledge evidence is verified against — a reader who selects a line here is selecting text the backend can find again. A viewer's own text layer would extract different characters and every capture from it would fail verification.

Pages come back whole, up to a character budget; next_page names where to resume so a long document loads in bounded chunks.