Uploading a file lets the model quote it, not guess it

Start here

A person uploads a contract or a research paper and asks a specific question about it, and the answer feels sharper and safer than the same question asked with nothing attached. The mechanism is called retrieval: instead of answering purely from what it learned in training, the model searches the text a person actually provided and builds its answer from passages found there. OpenAI's retrieval documentation describes this as semantic search, a technique surfacing similar results even when they match few or no keywords, then combined with a model to synthesize a response.

What the documents say

The vendors document real limits alongside the benefit. OpenAI's file uploads FAQ states that outside ChatGPT Enterprise's visual retrieval feature, other plans only support text-based retrieval, so ChatGPT extracts digital text from a file and discards any images in it; a chart or scanned page can be invisible to the model even though the upload succeeded. The retrieval guide separately documents caps: a maximum file size, a token limit searched per file, and a default of 10 search results per query, rising to 50. Google's Gemini Notebook overview describes a product built around this mechanism, giving grounded information with in-line citations, and stating that when a question cannot be answered from uploaded material, that refusal is by design, not a bug. The page states plainly that it can make mistakes and its answers do not reflect Google's views, so grounding reduces but does not remove the risk of an incorrect answer.

Check this

A reader relying on an uploaded document can test whether an answer is genuinely grounded by asking for the exact passage it drew from, then opening the document to confirm that quote appears there. If a document contains scanned pages or charts, the file-uploads limitation is the first thing to check: text-based retrieval on most plans discards images, so a chart question may be answered from a caption rather than the chart itself.

What holds and what fails

Retrieval genuinely changes what an answer is based on, and grounding in the user's own material, checkable through citations, holds as far as the citation actually points to real supporting text. It fails wherever the document is not fully readable by the system, as with images on non-Enterprise ChatGPT plans, or wherever a question falls outside every uploaded source, which both vendors document as a case meant to be declined, not guessed at. It is an editorial point, not a vendor claim, that a citation pointing to real text is still worth opening rather than trusted on sight.

• Ask for a direct quote from the uploaded document and check it actually appears there.
• Check whether a scanned image or chart in a file is likely to be readable by the retrieval mode in use.
• Expect an honest not-in-the-source answer when uploaded material does not cover a question, rather than a guess.

Retrieval is a narrower promise than it sounds: not that the model knows everything in a file, but that it can search and quote the parts it can actually read. That gap is where uploads still go wrong.