Hallucination Detection

• Key points are in the visual diagram above.

Hallucinations in support, legal, or medical agents cause real harm. Detection layers gate outputs or trigger retrieval retries.

Hallucination detection is a fact-checker with highlighter — every claim must be tied to a source line or flagged red.

Scenario

Agent says '30-day return policy' but retrieved doc says 14 days — NLI scorer flags contradiction; agent regenerates with correction.

Solution

In Eval Engineering & Observability, apply Hallucination Detection to this scenario: Agent says '30-day return policy' but retrieved doc says 14 days — NLI scorer flags contradiction; agent regenerates with correction. Identify the inputs, run the technique, validate the output, and note one thing you would monitor in production.

Do this before moving to the next module — reading alone is not enough.

Open the Code Walkthrough below and run it locally. Change one parameter related to Hallucination Detection (e.g. model, temperature, top_k, or tool name), observe the difference in output, and write 2–3 sentences explaining what changed.

Highlighted lines show where Hallucination Detection happens in the code.

1def groundedness_score(answer, context):  # define a reusable function2  prompt = f"Context: {context}\nAnswer: {answer}\nList unsupported claims."3  return judge_llm.invoke(prompt)  # returns list of unsupported spans

• Treating Hallucination Detection as a black box without evaluation
• Ignoring cost and latency in production
• Skipping error handling for hallucination detection

Quick recap — the most important points from this module.