Verification Prompts

Introduction

A verification prompt double-checks an output against the source data or a rule set. It's the cheap way to catch hallucinations before they reach users.

Beginner analogy: Like a second pair of eyes on a tax return. One pass writes, the other checks.

Understanding the topic

Core concepts to understand:

• After main answer, run: 'Does each claim in the answer appear in the source?'
• Output: pass / fail + list of unsupported claims.
• Use a cheaper model for verification when possible.
• Failed verification triggers retry or fallback.

Syntax reference

Visual workflow / architecture:

Generate answer
   ▼
Verify each claim against source
   ▼
PASS  →  return answer
FAIL  →  retry / mark uncertain

Real-world use

Perplexity uses verification-style prompts to ensure cited claims appear in cited sources.

Best practices

• Verify claim-by-claim, not paragraph-by-paragraph.
• Use a smaller, cheaper model for verification.
• Log verification failures for prompt improvement.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. Why verify after generating?
• Q2. How catch hallucinations programmatically?
• Q3. Cost of verification?
• Q4. What does the fallback look like?