Reducing Hallucinations

Introduction

A hallucination is when the LLM confidently states something false. Reducing them is half of production prompt engineering.

Beginner analogy: Like a confident toddler — sounds sure, often wrong. Engineering removes the false confidence.

Understanding the topic

Core concepts to understand:

• Ground answers in retrieved context (RAG).
• Ask for sources or 'I don't know' as a valid answer.
• Lower temperature on factual tasks (0 or 0.1).
• Verify with a second prompt.
• Use a smaller, calibrated model for fact retrieval.

Syntax reference

Visual workflow / architecture:

User question
   ▼
Retrieve relevant docs (RAG)
   ▼
"Answer ONLY from the context.
 If not in context, say 'I don't know.'"
   ▼
Generate
   ▼
Verify claims appear in context

Real-world use

Perplexity, Glean, Hebbia — all built on hallucination-reducing prompt patterns + RAG.

Best practices

• Allow 'I don't know' as a first-class answer.
• Always show sources to the user.
• Hard-fail downstream code on unverifiable claims.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. What is a hallucination?
• Q2. How reduce hallucinations programmatically?
• Q3. Why does RAG help?
• Q4. Role of temperature?