Hallucination Reduction Tasks

Introduction

Pick a hallucination-prone prompt and reduce it to <5% via grounding, verification, and structured output.

Beginner analogy: Like bug fixing — measure first, fix systematically.

Understanding the topic

Core concepts to understand:

• Add RAG context.
• Add 'cite source for every claim'.
• Verify with a second LLM call.
• Block unverifiable claims downstream.
• Measure on a 100-case eval set.

Syntax reference

Visual workflow / architecture:

Baseline rate → Add RAG → Add verify → Add downstream block → Re-measure

Real-world use

Common interview project — bring numbers and screenshots.

Best practices

• Build a 100-case eval first.
• Layer fixes one at a time.
• Track rate per fix.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. Hallucination measurement design?
• Q2. Why layer fixes?
• Q3. Downstream block — how?
• Q4. Reasonable target rate?