How assistants read files

AI assistants that accept uploads generally do one of two things with a document:

  • Read its text layer. For a text-based PDF or Word file, the extracted text is what the model works with. Layout, columns and tables have to be inferred from that text, which is where structure can get lost.
  • Look at it as an image. For scans and photos, the system either runs text recognition or lets a vision model read the page image directly. Quality then depends heavily on resolution, lighting and orientation.

Many services also have limits on file size, number of pages or the amount of text they can consider at once. Content beyond those limits may be skipped or summarised less carefully, often without a clear warning.

Step 1: give it text where you can

If the document exists as a Word file, web page or text-based PDF, use that version rather than a scan. If you only have a scan, run OCR in an app that offers it, then check that you can select and search the text. Our guide to how OCR works explains why scans need this step and which errors to watch for.

To turn a text-based PDF into an editable document first, use PDF to Word; it uploads the file to the Convertios server, converts it with pdf2docx and deletes temporary files after the request. It does not perform OCR.

Step 2: send only what is relevant

Long documents dilute attention and may hit limits. Extract the pages that answer your question with Split PDF, which runs in your browser. Asking about clause 14 of a contract? Send the pages around clause 14 and the definitions section, not the full 80 pages. This also reduces how much personal information you share; see our privacy checklist for AI tools.

Step 3: fix orientation and readability

  • Rotate sideways pages with Rotate PDF so text lines run the right way.
  • Retake blurry phone photos in good light, with the page filling the frame.
  • For a dense table, consider a crop of just the table at good resolution.

Step 4: tell the assistant what it is looking at

A short instruction helps the model use the document correctly:

  • "This is a scanned invoice; list the line items and the total."
  • "Pages 3–5 of a rental contract follow. Explain the notice period and quote the clause you rely on."
  • "The table has three columns: date, description, amount."

Asking for quotations or page references makes answers easier to verify.

Step 5: verify the answer

AI assistants can produce confident, fluent mistakes, especially with numbers in tables, small print and handwriting. For anything important:

  1. Check every figure the answer depends on against the original.
  2. Look up quoted clauses or sentences in the document.
  3. Be wary of summaries that contain information you cannot find in the file.

Quick checklist

  • Text-based version or OCR'd scan, not a blurry photo.
  • Only the relevant pages, upright and readable.
  • A clear instruction about what the document is.
  • Answers checked against the source, especially numbers.