Answer Grounding Techniques

"Grounding" is the process of ensuring that every claim in the LLM's response can be traced back to a specific piece of evidence in the retrieved context.

Strategy 1: The "I Don't Know" Fallback

This is the most important grounding rule.

Instruction: If the answer is not explicitly stated in the provided documents, state "Information not found." 
Do NOT use your own knowledge to supplement the answer.

Strategy 2: Quote Extraction

Ask the model to provide the exact quote from the document before it gives its answer. This forces the model to "read" the text carefully.

Step 1: Find the most relevant quote.
Step 2: Explain the quote in the context of the user query.

Strategy 3: NLI (Natural Language Inference)

Use a second, smaller model to verify the output.

• Premise: The retrieved context.
• Hypothesis: The model's answer.
• Verification: Does the premise entail the hypothesis?

Strategy 4: Self-Correction Loop

Ask the model to review its own work:

"Read your response above. Does it contain any information not found in the <context> tags? If so, remove it and rewrite."

Evaluation Metrics (RAGAS)

• Faithfulness: How much of the answer is derived directly from the context?
• Answer Relevance: Does the answer actually address the query?

Exercises

1. Compare two prompts: one with "Only use the context" and one without. Which one produces fewer creative details?
2. What are the downsides of forcing the model to only use provided context? (Hint: Does it become too rigid?)
3. How would you "ground" a response based on an image?