Identifying Agent Hallucinations

Agentic Output and Identifying Hallucinations

Context

Agentic output is here. It does a wonderful job of generating output. It combines the LLM's existing knowledge with the additional resources you provide to generate the outcomes.

---

Hallucinations

Hallucinations are statements in the output that are irrelevant or lack supporting context.

---

Detecting and Identifying Hallucinations

The great thing about agents is they can tell you when they are inferring in their response. You build it into the agent's system prompt. The identification of inferred output can be as explicit as you want it. Telling the agent that any time you output inferred text, tag it, change the font to italics, whatever makes sense for your situation. The goal is the text is clearly identified as inferred output. All inferred output is not bad. Does the output have context that supports the reasoning, or is it a wild guess/hallucination? That's its job. Provide us insight that we don't have from the data provided, but it needs to be reviewed.

Need Help with AI in Your Business?

Contact the Sales Team