Knowledge & tools

Verify Claim tool

A built-in tool that independently checks a single factual claim against the bot's real sources before the agent commits to it. Where confidence scoring and response validation grade an answer after it's written, Verify Claim lets the agent check a fact before it asserts it. The judgment is made by a separate evaluator, not the agent's own reasoning, so it can catch the agent's own mistakes.

This is the "prevent" layer of the bot's response-quality features. It pairs well with confidence scoring and response validation (which detect problems in a finished answer) and with the per-bot Quality review (where you see where a bot is weak over time).

Assign it

Verify Claim (type verify_claim) is a credential-free built-in, so it's immediately assignable with no config form. Turn it on per bot from the bot's Tools tab, exactly like search_knowledge or search_memory. There's nothing to configure.

Parameters

These show as the tool-call arguments on the tool card:

claim (string, required) - the single factual statement to verify. One claim, not a paragraph.
evidence (string, optional) - evidence the agent already gathered that it wants weighed in the check.

Result

The tool card's result preview shows a JSON verdict:

{
  "verdict": "supported | contradicted | unsupported | unverifiable",
  "explanation": "one sentence naming the deciding evidence",
  "evidence": ["the exact supporting or contradicting snippet(s)"],
  "sources_checked": [
    { "source": "knowledge_base", "hits": 3 },
    { "source": "web",            "hits": 1 },
    { "source": "tool_results",   "hits": 2 },
    { "source": "agent_supplied", "hits": 1 }
  ]
}

Verdict	Meaning
`supported`	A source states or clearly implies the claim.
`contradicted`	A source conflicts with the claim.
`unsupported`	Sources exist, but none are relevant to the claim either way.
`unverifiable`	No sources were available at all (no knowledge base, no web tool, no tool results, no supplied evidence). This is honest by design, not an error - the tool never fabricates a verdict, and `sources_checked` is empty.

sources_checked lists exactly which sources the check consulted and how many hits each gave; source is one of knowledge_base, web, tool_results, or agent_supplied. In the console you can read it at a glance as a "checked: KB, web" line alongside the verdict.

What it checks against

The tool automatically uses every source the bot actually has - you don't wire anything up:

Knowledge base - always, when the bot has one.
Web - only when the bot has a web-search tool assigned (Brave, Tavily, or DuckDuckGo); it reuses that tool.
This turn's tool results - the outputs of other tools the agent already ran.
Agent-supplied evidence - whatever was passed in the evidence parameter.

A richly-configured bot gets stronger verification; a bare bot with no sources returns unverifiable. To make Verify Claim useful, give the bot a knowledge base and/or a web-search tool.

When the agent uses it

The tool's own description steers the model to use it judiciously - for high-stakes or uncertain factual claims (figures, dates, policies, names, product details), not for every sentence. Each call does an evidence retrieval plus one independent judgment call, so it's heavier and slower than a plain search; the agent weighs that before invoking it. Its LLM cost is metered on your workspace usage like any other tool.

How it appears

Verify Claim renders as the standard tool card - the arguments, then the result above, keyed to the call. The internal source lookups (knowledge base, web, tool results) don't appear as separate cards; they're summarized in sources_checked on the one card. Like every tool, its output is operator-only: verdicts are never shown to end customers in the chat widget or messaging channels.

Classification & lifecycle

This tool is read-only - it's not mutating, destructive, or egress, so it's never blocked in autonomous runs and needs no confirmation. It's post-call-hook eligible (runs safely outside the LLM turn).