ASAmol Shukla
55 mins lesson duration•10 mins read

Lesson 9: Detecting & Reducing Hallucinations

Grounding, RAG, citations, self-checking prompts, and workflow design — the practical toolkit for trustworthy LLM output.

The Verification Toolkit

You cannot eliminate hallucinations — but you can architect systems that catch them. These techniques stack; use more of them for higher-stakes tasks.

1. Grounding (Give the Model Facts)

Never ask the model to recall facts you already have. Put the source material in the prompt:

Answer ONLY using the document below. If the answer is not in the
document, say "Not found in the provided document."

DOCUMENT:
[your data here]

2. RAG (Retrieval-Augmented Generation)

For large knowledge bases, retrieve the relevant chunks (Lesson 5, 13) and ground the answer in them. This is the professional standard for chatbots over your own data.

3. Citations & Traceability

Require the model to cite which part of the provided text supports each claim:

"Answer with inline references like [section 2.3] or [doc 1, para 4]. If a claim is unsupported, label it UNSUPPORTED."

Then a human (or a checker script) can verify.

4. Structured Self-Check Prompts

  • Ask the model to quote the evidence before answering.
  • Ask it to list assumptions it made.
  • Give it the option to say "unknown" — explicitly rewarding honesty over completion.
  • Run a second pass: "Review your previous answer. Which claims are not supported by the source? Revise."

5. Post-Processing Checks

Check Catches
Regex/JSON schema validation Format hallucinations
Known-entity whitelist Invented names/IDs
Citation-to-source matching Fabricated references
Round-trip checks (summarize → re-read) Drift
Human review for high stakes Everything else

6. Calibrate the System, Not the Model

Lever Effect
Lower temperature (0–0.3) Fewer random inventions
Smaller, focused prompts Less room to drift
Grounding + RAG Eliminates the "recall from nowhere" problem
Allowed answer: "I don't know" Removes the pressure to invent
Human-in-the-loop review Catch what automation misses

A Simple Grounding Prompt Template

You are a fact-checked assistant. Rules:
1. Answer ONLY from the CONTEXT below.
2. Quote the relevant CONTEXT line before each claim.
3. If CONTEXT does not contain the answer, reply exactly:
   "I cannot answer from the provided context."

CONTEXT:
{context}

Key Takeaways

  • Ground every important answer in source material — never free-recall.
  • Give the model permission to say "I don't know".
  • Require citations and add automated checks on top.
  • Stack techniques: grounding + RAG + verification + human review.

Next up: Agentic loops — turning LLMs from single-shot answerers into multi-step workers.

Interactive Lesson Code Snippet
# Reducing hallucinations: verify model answers against a source of truth
source = {
    "Paris": "Capital of France since 508 AD.",
    "Python": "Created by Guido van Rossum in 1991.",
}

def verify_answer(model_answer):
    for fact, truth in source.items():
        if fact in model_answer:
            return f"Verified: {truth}"
    return "Unverified - please check the source or add citations."

print(verify_answer("The capital of France is Paris."))
print(verify_answer("The moon is made of cheese."))

# A grounded generation prompt: the model may only use CONTEXT
def grounded_answer(question, context):
    # Simple relevance check: does the context mention the topic?
    topic = question.split()[-1].strip("?")
    if topic.lower() in context.lower():
        return f"Based on context: {context}"
    return "I cannot answer from the provided context."

print(grounded_answer("Who created Python?", "Python was created by Guido van Rossum in 1991."))
print(grounded_answer("What is the price of tea?", "Python was created by Guido van Rossum in 1991."))
Language: python

Code Visualization Tips

  • Draw the 'grounding sandwich': CONTEXT above and below the answer — the model reads facts, then writes within them.
  • Diagram the RAG verification loop: Answer → Cite → Check against source → Pass/Reject.
  • Make a checklist poster of the 6 levers: grounding, RAG, citations, self-check, post-checks, human review.

Professional Tips & Tricks

  • The phrase 'answer only from the context' alone cuts hallucinations dramatically — make it your default.
  • Add 'quote the evidence first' for any answer that will be shared externally.
  • For numbers: 'if the number is not in the source, say UNKNOWN' — never let the model estimate.

Python Code Judge & Practice Arena

LeetCode Style

Run real Python 3.12 WebAssembly code directly in your browser against automated test suites.

Solved:0 / 1
0 / 20 XP
Challenges:
Problem 1 of 1

Write a Grounding Prompt

Medium+20 XP
Write a system prompt that forces a support chatbot to answer only from the company FAQ, quote the FAQ line, and say 'I don't know' otherwise.
main.pyPython 3.12 (WASM)
1
2
3
4
5
6
7
8
9
10
11
12
Press Run Code to test or Submit to verify test cases

Up next · Continue learning

What Is an Agentic Loop?

One prompt = one answer. An agentic loop = the model plans, acts, observes, and repeats until the job is done.

10 mins read50 mins
Start next lesson
Why LLMs Hallucinate What Is an Agentic Loop?
Made withbyAmol Shukla·amolshukla.online