Day 19 — Stopping Hallucination

What Is Hallucination

🛡️

When AI Confidently Makes Things Up

Hallucination occurs when an LLM generates facts that sound plausible but are not in the retrieved context — or are simply wrong. LLMs are probabilistic: when they do not have the right answer, they still produce the most statistically likely next words. The result is confident, fluent, wrong information.

⚠️ Hallucination is dangerous because the LLM presents fabricated information with the same confidence as verified facts.

Root Causes

🔍

Why Hallucination Happens in RAG

✕Irrelevant documents retrieved — LLM never received useful context to work from

✕No strict prompt grounding — LLM is free to go beyond the provided context

✕Missing information — Topic is not in your documents — LLM fills the gap with guesses

✕No verification step — Generated answer is never checked against the source

Layer 1 Prevention

🏛️

Strict Prompt Grounding

The most effective and simplest technique — tell the LLM explicitly to answer only from the provided context. This one instruction dramatically reduces hallucination in most RAG systems.

✓"Answer ONLY from the provided context below"

✓"If the context is insufficient, say: I don't know"

✓"Do not use your general knowledge or make assumptions"

✓"Do not state anything that is not explicitly in the context"

Layer 2 Prevention

🔒

Is Supported Verification Check

After generation, verify the answer against the retrieved context. An LLM evaluator reads both the answer and the source documents and assigns a support level — fully supported, partially supported, or not supported.

✓Fully Supported — Every claim came from the retrieved documents — safe to return

→Partially Supported — Some claims are fabricated — revise the answer and retry

✕Not Supported — Pure hallucination — reject, retry, or return I don't know

Layer 3 Prevention

🔗

Better Retrieval = Less Hallucination

If retrieval is poor, no prompt instruction will save the answer. Better retrieval means the LLM always has accurate, relevant context to draw from — leaving no gaps it needs to fill with guesses.

💡 Anti-hallucination stack: Better Retrieval + Strict Prompt Grounding + Is Supported Verification = Near-zero hallucination in production.

✦

Stopping hallucination requires a multi-layer approach. Layer 1 — strict prompt grounding forces the LLM to use only provided context. Layer 2 — the Is Supported check verifies every claim post-generation. Layer 3 — better retrieval ensures relevant context exists. Three layers together produce a production RAG system where hallucination is the exception, not the rule.

← Day 18 Day 19 / 20 Day 20 →