Files

A short document goes to the assistant whole. Anything up to 40,000 characters, with up to 60,000 across the files of one message, is placed in the turn as if you had pasted it, so a brief or a spec is followed in full. A longer document is chunked, indexed and searched, so a 200 page PDF costs about as much context as the few paragraphs that turn out to matter. The assistant can also read one through in order, page by page, when it needs all of it. Send the whole thing. Pasted text stays in your message up to 100,000 characters. Only a longer paste is offered as an attachment instead.

​

What you can attach

Plain text and markdown, CSV and JSON, source files in most common languages, PDF, Word, Excel and PowerPoint. Images are different. They are not indexed, they go straight to the model, so they need a model that can see. A model without vision says so rather than failing quietly. Animated GIFs are rejected.

​

How search works

Each document is split into overlapping chunks of about 200 tokens. Two indexes are built: a keyword index and a vector index. A search runs both, takes the top 30 from each, merges them by rank, and returns the best five. That combination is why an exact identifier and a vaguely worded question both find the right passage. Keyword search catches the first, meaning search catches the second.

​

Identical files are free

A file is identified by its content. Attaching the same document again in another chat reuses the existing index and does not count against your upload quota.

​

Upload quotas

Images do not count against these.