How file uploads work
Larger files are indexed, not read in full, the model sees only the most relevant passages, so specific questions get far better results than vague ones.
Not all uploads work the same way
When you upload a file to ChatGPT, you might assume the model reads every word before answering. For small files, that is roughly what happens, the whole document can fit in the context window (the amount of text the model can read at once) alongside your conversation, and the model sees it in full.
For larger files, something different happens: the file is indexed. The content is split into chunks and stored in a way that makes it searchable. When you ask a question, the system retrieves only the chunks most likely to be relevant, and adds those to the context. The model then answers from those retrieved passages, not from the full document.
This retrieval approach lets ChatGPT handle files much larger than would fit in the context window. But it also means the model is working from a selected excerpt. Broad or vague questions can miss information that is technically in the file, because the retrieval step did not pull the right chunks.
Select a numbered hotspot to see what occupies that part of the diagram.
Attach files, change the answer
Toggle the knowledge files below. The model can only use what's attached, watch the answer change as you attach and detach the pricing sheet.
What the AI can see
No files attached, the model sees only your question.
Question: What does the Pro plan cost per month?
Answer: I don't have that information. None of the attached files mention pricing, attach pricing-sheet.xlsx to answer this.
The model only knows what's in its context, attaching a file puts it within reach; detaching one takes it away, even if the model answered from it a moment ago.
Question 1 of 4
For a large uploaded file, what does ChatGPT actually read when answering a question?