Validating Responses

RAG reduces hallucinations, but doesn't eliminate them. The model might still try to be "helpful" by inventing facts if the retrieved chunks are irrelevant.

The "I Don't Know" Rule

Add this to your System Instruction:

"If the answer is not contained in the provided context, state 'I do not have enough information'. Do not make up an answer."

Citation Requirement

Force the model to cite the chunk ID.

"Every sentence must end with a citation like [Source: Doc 1]."

If the response lacks citations, treat it as suspect.

Evaluation (RAGAS)

Use a framework like RAGAS (Retrieval Augmented Generation Assessment).

• Faithfulness: Does the answer match the context?
• Relevance: Did the retrieval step find useful chunks?

Summary

RAG without validation is dangerous. Use prompt constraints and automated evals to keep the system honest.

Module 8 Complete! You have built a knowledge engine. In Module 9, we automate it: Workflows in AI Studio.