Where an answer came from
Every answer is produced by tools that ran against your workspace, and the status line's folded detail lists them: which tools, in what order, how many steps. Answers grounded in your own material name what they were drawn from — a thread, a file, a knowledge source, a memory. An answer that can name nothing is telling you something, and it is not that it is confident.
Two kinds of evidence#
The first is procedural: what the Assistant did. That lives in the status line — Understanding… while it works, then Done · N steps, with the detail folded away until you open it. Unfolding shows the tools that ran, which is the difference between an answer assembled from four lookups and one that consulted nothing.
The second is substantive: what the answer rests on. A claim about a customer should point at the conversation, follow-up or memory it came from; a claim about your business should point at the knowledge source or fact behind it. Those are things you can open and read for yourself, which is the only definition of a citation worth having.
| Claim about | Cited as | Where you can check it |
|---|---|---|
| A conversation | The thread and the message in it | Conversations, #/inbox |
| A commitment | The follow-up, with its date and reason | Follow-ups, #/follow-ups |
| A person's history | The timeline entry on that channel | The relationship, #/relationships |
| A business statement | The knowledge source, or a fact drawn from it | Knowledge, #/knowledge |
| Something Connect was told | The memory row and the tier that owns it | The memory viewer on that contact or channel |
| A file | The file and its version | Files, and the file's own history |
Reading the step detail#
Open the folded detail on the status line after an answer you want to check.
Result You see the tools that ran. Names beginning
list_,get_,read_,search_anddata_queryare reads; anything else was an action you confirmed.Compare the count against the answer's confidence.
Result A definite claim built on one step deserves more scrutiny than a hedged one built on six. Steps are not accuracy, but a claim with no lookup behind it has nothing behind it.
Open the source itself when the answer will decide something.
Result The cited thread, follow-up or knowledge source opens in the app. The Assistant is a faster way to find the evidence, not a substitute for looking at it.
When it cannot cite anything#
This happens, and it is the most useful moment on the page. There are three ordinary causes, and they call for different responses.
- Nothing has been recorded
- The honest answer is 'nothing in this workspace says so'. That is information: it usually means a conversation happened somewhere Connect cannot see, and the fix is to write it down as memory or knowledge rather than to ask again.
- The material exists but is not grounded
- Knowledge sources are what answers are grounded in. A PDF nobody has added as a source is not knowledge, however long it has been in the files list. Add it — see Knowledge in Connect.
- The claim is general rather than about you
- An answer about how email works has no workspace source and should not invent one. Take it as background, and treat any claim about *your* configuration as needing a citation.
Grounding, honestly described#
Connect grounds answers in the knowledge a workspace supplies and in the facts drawn from it, with a deliberate budget on how much reaches the model at once. That budget is why a grounded answer cites two or three sources rather than twenty: the material is selected, not poured in.
The corpus does not describe that selection as a vector search, because claiming machinery Connect does not run would be exactly the kind of unverifiable statement this page is about. Grounding an answer without pretending to have a vector database sets out what is actually done.
Questions#
Why does the same question sometimes cite different sources?
Because your workspace changed, or because a different selection of material fit the budget this time. If two answers disagree, that is a signal in itself — When sources disagree covers what Connect surfaces and who decides.
Can I get a citation for an answer about a file?
Yes. An answer built from a file names the file and the version it read, which matters because Connect creates and modifies files as new versions rather than overwriting them. An answer about version three is not an answer about version four.
Does the decision log hold the evidence too?
The decision log holds what was decided, under which rule, and what happened — including refusals. It is the accountability trail for actions rather than a transcript of the reasoning behind an answer.