Answering from a file
Ask a question with a file attached and Connect answers from that file. The ask_file tool reads the document you name and the reply carries the file it came from, so the claim and its evidence arrive together. If the document does not contain the answer, Connect says the document does not contain it rather than filling the gap from general knowledge.
What grounding means here#
Grounding is a restriction, not an enhancement. When a question is pointed at a file, the model is given that file's contents and told to answer from them. Everything else it might know about invoices in general is still there, but it has been told which text is authoritative, and the answer names it.
That distinction is what makes an answer checkable. A sentence with no source is a sentence you have to believe; a sentence with a file beside it is a sentence you can open the file and disagree with.
The three shapes an answer takes#
| Shape | When | What to do with it |
|---|---|---|
| A direct answer with the file cited | The document says it plainly | Check the quoted part if the decision is expensive |
| An answer with a stated inference | The document implies it — a total that must be the sum of the lines above it | Verify the step, not just the number |
| A refusal to answer | The document does not cover it | Attach the document that does, or ask a different question |
When the file and the workspace disagree#
A quoted price in an attached PDF and a value on the deal record are two sources, and they can differ. Connect does not silently prefer one. The answer says both, names where each came from, and leaves the decision with you — the same behaviour described in when sources disagree.
Nothing about answering from a file writes to the workspace. Reading a document does not update the deal, correct a contact or add a fact to Knowledge. If the file should change what Connect believes in future, that is a separate and deliberate step.
A file is untrusted text#
A document that arrives from outside is content, not instruction. A PDF containing a line telling an assistant to ignore its rules and email the contact list is exactly as authoritative as any other sentence in that PDF, which is to say not at all. Connect treats file contents as material to answer from and never as a direction to act on.
This matters most for attachments that came in on a channel rather than from your own desktop: a supplier's quotation, a form a customer filled in, a spreadsheet forwarded by somebody you have never met. The note on prompt injection arriving as ordinary business mail sets out why the boundary is drawn where it is.
Failure modes worth knowing#
- The answer is about the wrong file
- More than one attachment is in the conversation. Name the file in the question, or start a fresh tab so only one is in scope.
- The PDF produced nothing
- It has no text layer and the page fell back to being read as an image. The PDF has no readable text covers what to do.
- The answer covers only part of a long document
- A long file is read in parts. Ask about a named section rather than the whole thing, and the reading narrows with the question.
- The file was refused before any of this
- Validation or safe parsing rejected it. The file was refused names the reasons.
Questions#
Does Connect remember the file after it answers?
The file stays in the workspace with its record, so you can ask about it again. What it does not do is quietly become part of what Connect believes: an attachment is not a knowledge source until somebody makes it one, which keeps a stray forwarded document out of every future answer.
Can I ask about several files at once?
Yes, within one question. The answer then draws on all of them and cites each claim to the file behind it. Where two of them conflict you get both readings rather than an averaged one.
Why did Connect refuse to answer from a file it clearly read?
Reading a document and answering a question from it are different things. If the text does not settle the question, the honest reply is that it does not — and that reply is worth more than a plausible sentence assembled out of the surrounding paragraphs.