GeminiPrompting mistakesLevel 2: Context

Why Did Gemini Miss Part of Your File?

Usually the file was too big for your plan's context window, or you asked for too many things at once. Google says exceeding the window can cause responses to miss details spread through large files. A successful upload doesn't mean Gemini took in every page, so ask it to cite where each answer came from.

Prompt to copy
Before answering, tell me what you can see in the attached [file]: total pages or sections, the first and last headings, and any part you couldn't read.

Then answer one question: [one specific question].

For each claim, give the page or section. Quote the sentence you relied on. If the answer isn't in the file, say "Not found" instead of inferring it.
  • 100 MBMost files can upload
  • 10Files per prompt
  • 32kWindow with no plan
  • 1MWindow on Pro and Ultra
Why a file can upload and still get missed: the upload limit is far larger than the window. Source: Google's upload and limits pages, checked November 6, 2026.
  1. 1Does it fit?32k to 1M tokens
  2. 2Is it readable?rotated, not blurry
  3. 3One ask at a time?accuracy drops
  4. 4Ask where it found itpage or section
Check in this order. Rule of thumb, based on Google's limits and upload pages.

Try it now

  1. 1Click Add files, then Upload Files, and attach the file that gave you trouble.
  2. 2Paste the prompt above. Check that the last heading it reports really is the last one.
  3. 3If it can't see the end, split the file and upload only the part you need. Ask one question per message.
  4. 4You'll know it worked when each quote is in the file, on the page Gemini named.

Use it when

When a summary skips a section, Gemini says something isn't in the document when it is, or the answer draws only from the first part of a long file. It's a good check on any file where you'll rely on the answer.

Skip it when

If Gemini says Gemini failed to analyze your file, that's a different problem, and Google's fix is to re-upload. And if the message is You've reached your limit for chats with files, that's a usage limit, so splitting the file won't help.

Why it works

Upload limits and context windows are separate things. Google's upload page allows up to 10 files per prompt, with most files up to 100 MB. The context window is much smaller on lower plans: 32k tokens with no plan and 128k on AI Plus. A 100 MB file can pass the upload check and still far exceed the window. You may see Your uploads may be too large for the best results, and Google's fix is smaller files or a larger window. Separately, its API docs say accuracy drops with many facts to find in one request, and scans read worse if they're rotated or blurry.

Example

Say you upload a 300-page employee handbook on a free account and ask about parental leave. The answer quotes an early chapter and says nothing about the section on page 240. At 32k tokens, that's expected, since the whole handbook doesn't fit. Export pages 230 to 250 as a new PDF, upload that, and ask again. On AI Pro, Google says the 1 million token window holds roughly 1,500 pages of text, so the whole book would be a more realistic ask.

Going further

Plain text and PDFs behave differently. Google's document docs say the model reads PDFs with native vision, including charts and tables, while formats like TXT, Markdown and HTML are read as text only, so charts and formatting are lost. If your answer needs a chart, send the PDF rather than a text export. If you need 20 specific facts from a long file, ask for them across several messages. Google's long context page says a single fact is what it extracts most accurately. Deep Think has a 192 thousand token window, so don't choose it for size.

Common mistakes

  • Assuming an accepted upload was read in full. The upload limit and the context window are different limits.
  • Asking for "everything important" in a 300-page file. Ask specific questions, and ask for citations.
  • Uploading a sideways scan. Google's docs say to rotate pages to the correct orientation and to avoid blurry ones.
  • Exporting a report to plain text to make it smaller, and losing the charts that held the answer.

Sources

Checked 6 November 2026. Products change often, so confirm details with the vendor.

Related

Beyond the course