# Rating an answer Rating an answer records your judgement against that answer and the conversation it belongs to. It is a signal, not a control: no model is trained on it, and the next answer is not different because you rated the last one. When you want the behaviour itself to change, correct the fact in Memory or write a standing instruction. Those two do change what happens next. - **Status:** Foundation - **Audience:** both - **In the app:** #/home - **Last verified:** 2026-09-10 - **Canonical:** https://connectbyjbrh.com/docs/assistant/feedback/ ## What a rating captures A rating attaches to one answer, in one tab, in one conversation. It carries the judgement, the answer it refers to, and the workspace it happened in. It does not copy the answer text anywhere new, and it does not travel outside the workspace it was given in — a rating is workspace data under the same row-level security as everything else. The rating worth giving is the one where the answer was wrong for a reason a reader could name: it quoted a record that has since changed, it read a file that turned out to be the older version, it answered about the wrong person. Those are the reports that let a fault be found. A rating with no reason is a mood, and there is little to be done with one. ## The honest limit This is a foundation capability rather than a finished loop, and the narrower part is what happens afterwards. The rating is stored; there is no documented mechanism by which it alters a later answer, and Connect does not train a model on your workspace data. Anyone who tells you that thumbing an answer down teaches the Assistant is describing a different product. > **Careful** Do not use a rating as the way to correct a fact. A wrong price, a wrong contact or a wrong policy left in Memory stays wrong, and the next person who asks gets the same wrong answer with the same confidence. ## What does change the next answer 1. Say the correction to the Assistant in the same conversation — that is out of date, the delivery window is now two working days. - Result: A `remember` proposal appears, showing what would be stored and at which tier. Nothing is written until you confirm it. 2. Confirm it. - Result: The memory service writes the fact, and the next answer that needs it draws on the new one. The change is visible in [Memory](/docs/memory/) and carries who made it. 3. If the old fact is also in Knowledge, fix it there too. - Result: Grounded answers prefer Knowledge, so an unfixed source keeps producing the old sentence and the two then disagree — which becomes a [source conflict](/docs/assistant/source-conflict/), reported rather than quietly resolved. For behaviour rather than fact — tone, what to escalate, what never to promise — a standing instruction is the durable form, because it applies across channels rather than to one conversation. ## Rating a refusal A refusal is the answer worth rating most carefully, because there are four quite different reasons for one and only one of them is a fault. If the Assistant declined because the action sits outside its authority — pricing, or clearing a do-not-contact entry — the refusal is correct, and a low rating buries a real signal under a wrong one. [The Assistant refused to do it](/docs/troubleshooting/assistant-refused/) tells the four apart in about a minute. **Worth rating** — The answer was factually wrong, cited a record it did not read, or contradicted a fact in Memory. **Worth a correction instead** — The answer was right about what it knew, and what it knew is out of date. **Worth an administrator, not a rating** — The Assistant could not reach a record you can see on screen — an access or context fault. **Not worth a rating** — It refused an action it is not permitted to take. That refusal is the design working. ## Questions ### Does a low rating stop the Assistant repeating the answer? No. Ratings are recorded and are not wired to the model turn that follows. What stops a repeat is changing the thing the answer was drawn from: the memory, the knowledge source, or the record itself. ### Can anyone outside the workspace read my ratings? A rating is workspace data and stays inside the workspace under row-level security, like the conversation it belongs to. It is not published, not shared with another workspace, and not used as training material. ### Is there a written reason field? The sources this page is written from do not document one, so this page does not claim it. What is certain to be read is a correction made through `remember` or a standing instruction, because both change behaviour rather than only recording an opinion. ## Related - [Connect Assistant](https://connectbyjbrh.com/docs/assistant/) - [Correcting a fact without leaving the Assistant](https://connectbyjbrh.com/docs/assistant/assistant-memory/) - [Memory in Connect](https://connectbyjbrh.com/docs/memory/) - [When the Assistant is not sure](https://connectbyjbrh.com/docs/assistant/uncertainty/) - [Making a correction actually change behaviour](https://connectbyjbrh.com/research/memory-correction-that-sticks/) ## What this page is based on - `docs-source/sources/GENERAL.md` §8 — the Assistant and its tools - Connect capability registry (docs-source/facts.py) — `memory_editing`, FORBIDDEN_CLAIMS - `backend/app/connect_memory.py` — memory tiers and correction