Confidence thresholds determine how the platform categorizes each AI response into high, medium, or low confidence bands, and what minimum search relevance the orchestrator requires before it is willing to stream an answer at all. These bands drive escalation rules, dashboard metrics, and audit records — tuning them is one of the fastest ways to change how the AI assistant behaves without touching any content.

The Confidence Thresholds page combines the band slider with the Minimum Reranker Threshold control.

Confidence Bands

The upper panel displays a dual-handle slider ranging from 0% to 100%. Drag the handles to set two boundaries that split every scored response into three bands:

• High confidence — responses scoring above the upper threshold (default: 90%)
• Medium confidence — responses scoring between the two thresholds (default: 60%–90%)
• Low confidence — responses scoring below the lower threshold (default: below 60%)

The slider uses a color gradient — green for high, yellow for medium, and red for low — so you can see the relative size of each band at a glance.

Previewing the Impact

When you adjust either threshold, a reclassification preview appears below the slider. The preview analyzes the last 30 days of conversation data and shows:

• Current vs. proposed counts — how many conversations fall into each band under the current and proposed thresholds
• Reclassification impact — which conversations would move between bands (for example, from medium to low or from low to medium)

This helps you understand the effect of a change before you save it.

Minimum Reranker Threshold

The second panel on the page holds a single slider labelled Minimum Reranker Threshold. This control sets the minimum search relevance score that the orchestrator requires from your knowledge base before it will generate an answer at all. It operates earlier in the pipeline than the confidence bands — the reranker threshold is a gate that decides whether a response is worth producing, whereas the confidence bands classify responses that have already been produced.

Use the slider to tune how strict the knowledge base match must be. The orchestrator compares the highest reranker score from the retrieval step against this threshold:

• If the best match meets or exceeds the threshold — the orchestrator proceeds, streams an answer, and scores it for confidence as usual
• If every result falls below the threshold — the orchestrator aborts the generation step and returns a low-confidence status without producing an answer

A higher threshold (for example, 0.7) forces the assistant to stay silent on questions where no confidently-matching content exists, which is the right choice for compliance-sensitive tenants. A lower threshold (for example, 0.2) lets the assistant attempt an answer even when the knowledge base signal is weak, which is useful for tenants with a small knowledge base that still prefer a best-effort response over nothing at all.

Saving Threshold Changes

Field Reference

Related Pages

• Escalation Rules — define automated actions based on confidence bands
• Model Routing — configure the confidence assessment model used to score responses
• Dashboard — monitor confidence distributions and success rate trends
• Conversation Logs — review audit records with confidence scores