Sharing files puts them in the conversation, not a separate lookup
When you upload a document or image, its content becomes part of the conversation's context, the same context window everything else shares, rather than something Claude looks up on demand later. This is why it's worth being specific about what you actually want done with an uploaded file, rather than just attaching it and asking a vague question: "summarize the key risks in section 3" gets a far more useful response than "what do you think of this," because the model isn't guessing at your goal on top of processing the file.
For images specifically, Claude can describe what's in a photo, read text captured in an image, and answer questions about layout, charts, or diagrams shown in it. As with any output, treat a description of fine detail, small text, or an unusual chart as something worth double-checking against the original if it matters, the same way you'd double-check any other Claude output that needs to be precise.