Safety and guardrails

Safety is a read-only record of the guardrail decisions a knowledge base has made over the last 30 days. It shows you what was blocked or suppressed; the dials that decide how readily an answer is given at all live on Settings.

What you'll need

• The Knowledge Base module visible in your left-hand navigation.
• This permission: Knowledge Base · Kb — VIEW.
• A knowledge base people are already asking questions of — the feed only records decisions taken on real queries. Changing a strictness level additionally needs kb.kb.update.

If Knowledge Base isn't in your navigation, your role doesn't have visibility for it — ask a workspace administrator, or see Permissions and module visibility.

Steps

1. Open the knowledge base, then select Safety in its section navigation.

2. The tiles are Prompt-injection blocked, PII redactions, Low-groundedness suppressed and Abstentions. They tally only the rows loaded so far, so loading further pages raises them — treat them as a sample, not a total.

3. Each row carries Time, Event and Detail. Event is Injection blocked, PII redacted, Low groundedness or Abstention. Detail is the document id where there is one, and nothing more — never the question, the passage or the redacted value.

4. Where the decision was recorded against a query, the row offers Open trace. It opens that query’s retrieval trace, so you can see what was retrieved, used or filtered out.

5. Nothing on this screen is a switch. On Settings, set Relevance strictness and Groundedness strictness to Low (lenient), Medium or High (strict) — High (strict) declines more often, Low (lenient) answers more on weaker evidence. Q&A matching does the same for curated pairs. Select Save changes.

What success looks like

• The tiles and the table list this knowledge base's guardrail decisions from the last 30 days.
• Selecting Open trace lands you on the query behind a decision.
• After a strictness change, Settings saved successfully! appears and the next question asked uses the new level — no re-indexing is involved.

What the feed does not prove

An empty feed is not a clean bill of health. The screen says as much: An empty feed means no decisions have been logged in this window. The window is fixed at 30 days with no date filter, and injection blocking and abstention are not switched on or off anywhere in the knowledge base — the strictness levels are the only guardrail settings you control.

If something goes wrong

Symptom	Likely cause	What to do
No guardrail activity yet.	Nothing has been logged for this knowledge base in this window.	Expect this on a quiet knowledge base. Read it as "nothing recorded", not "nothing happened".
The tile counts keep rising as you read.	They tally the rows loaded, not the window.	Load every page before comparing figures.
Readers report too many "no answer" replies.	Groundedness strictness or Relevance strictness is stricter than the content supports.	Lower one level on Settings. Chat names the gate that stopped it — Stopped at: Relevance gate or Stopped at: Groundedness gate.
A row offers no Open trace.	That decision carries no query to open.	Find the query on Retrieval trace instead, using its time and result.
The pane is replaced by Access Forbidden.	Your role cannot read this knowledge base.	Every tab is visible to everyone, so the refusal arrives when the data loads. Ask an administrator to tick VIEW on Knowledge Base · Kb for your role.

Next

• Settings — where the strictness levels are saved.
• Retrieval traces — what a single query retrieved and why.
• Access — who can reach the knowledge base and its documents.