Confidence

When to useCopy Anchor Link

• The user will act on a claim the agent isn't fully sure of: a cause, a diagnosis, a figure pieced together from several sources.
• The sources disagree, are stale or thin, and the user should see that before relying on the answer.
• The agent could check further (fresher data, another source), but it costs time, so the user should decide.

When not to useCopy Anchor Link

• The answer is a plain fact the agent is sure of. Just answer.
• The uncertainty is in guesses the user could correct, such as party size or budget. Use Assumptions.
• The point is where each sentence came from. Use Citations.

How the agent should use itCopy Anchor Link

• Set level honestly from the evidence (high, medium, low) and hedge the claim to match. Put the hedge words in hedges; they're underlined in the claim.
• Keep reason to one short line, such as "2 sources agree, 1 is stale". It sits next to the level; the detail goes in evidence.
• Include evidence against whenever it exists. source and url come from tool results, never from memory.
• Give changeMind as concrete checks, especially at low and medium: what fresh data would move the level.
• Name the check in verifyLabel ("Check live usage", "Pull full logs"). Set canVerify: false when there's nothing more to check.
• Only send score if a tool or calibrated model produced it. The level is enough on its own.
• On verify, check further and show a new confidence card with the updated level. On accept, carry on with the answer.

InteractionCopy Anchor Link

• The level shows as three bars and a label; colour marks high and low only. A low level adds a "Treat this as a hypothesis" line and opens the evidence.
• Evidence is a disclosure listing what supports the claim, what cuts against it, and what would change the agent's mind.
• Verify and Accept are pill buttons; Verify is the primary one at low confidence. After either, the buttons are replaced by a record, which takes focus and is announced.
• While streaming, the claim, level and evidence arrive in turn and the buttons are disabled. A result from history restores the record without calling onResult.

InstallationCopy Anchor Link

npx shadcn@latest add https://your-site.com/r/confidence.json

UsageCopy Anchor Link

import { tool } from "ai"
import { confidenceTool } from "@/components/genui/confidence.schema"

const tools = {
  [confidenceTool.name]: tool({
    description: confidenceTool.description,
    inputSchema: confidenceTool.inputSchema,
  }),
}

import { Confidence } from "@/components/genui/confidence"

function ToolPart({ part, addToolResult }) {
  if (part.type !== "tool-show_confidence") return null

  return (
    <Confidence
      {...part.input}
      status={part.state === "input-streaming" ? "streaming" : undefined}
      result={part.state === "output-available" ? part.output : undefined}
      onResult={(result) =>
        addToolResult({
          tool: "show_confidence",
          toolCallId: part.toolCallId,
          output: result,
        })
      }
    />
  )
}

The result tells the agent what the user chose:

{ action: "verify" } // check further, then show an updated card
{ action: "accept" } // carry on with the answer

SchemaCopy Anchor Link

Host props: status (streaming, ready), onResult, result and className.

SourceCopy Anchor Link