How can I best reduce the risk of my agent showing hallucinations?

Large language models (LLMs) can unintentionally introduce extra details when they lack sufficient information.

To mitigate this issue, consider these strategies:

• Validate your agent before deployment: Use Monitoring and Evaluation tab to evaluate your agent against representative user scenarios and edge cases. Review the evaluation results to identify hallucinations, grounding issues, and other quality gaps, then refine your prompts and knowledge sources before promoting your agent to production.
• Supply the necessary knowledge: Integrate relevant sources such as business objects, RAG documents, external and internal sources added through APIs or Connectors, ensuring the LLM has access to accurate and up-to-date information.
• Use explicit prompt instructions: Clearly instruct the LLM not to add or infer information beyond what's provided, reducing the risk of embellishment or fabrication.