Bigger context is not the same as better signal

Gemini 2.5 has been trained to retrieve from very long contexts, but retrieval is only as good as what you put in. PDF source bloats the input with running heads, page-break artefacts, sidebar callouts, and OCR drift on scanned pages. The model then has to scan past all of it to find the actual claim you're asking about — and recall accuracy degrades faster than you'd expect once layout noise crosses ~30% of the input.

Converting to Markdown first removes the noise and turns ## headings into navigational anchors Gemini uses internally. The result on Gemini 2.5 Pro: faster responses, more confident citations, and the ability to fit several long documents in a single prompt without truncation.

Using Markdown with Google AI Studio and Vertex

Both AI Studio and Vertex AI accept .md attachments natively. In AI Studio, attach the Markdown and ask your question — Gemini treats it as a primary source. In Vertex, use the fileData field with mime type text/markdown; the API streams it as cached context, so you're not re-paying tokens on every call.