Hallucination Detection
A technology that checks whether responses generated by a large language model (LLM) are grounded in provided reference documents or contradict objective facts to identify potential misinformation.
Detailed explanation
Why It Matters in Tool Selection
Due to the non-deterministic nature of LLMs, incorrect responses can be generated at the service stage. When selecting an enterprise solution, you should verify whether it has a real-time detection workflow, how it calculates the alignment between the reference documents and the response, and whether the balance between detection accuracy and response latency meets your business requirements.
What to Check
- Check whether it quantifies and displays the level of alignment between the provided context and the response.
- Check whether it supports guardrail features that can filter or rewrite responses in real time post-generation.
- Check whether internal knowledge bases can be connected to verify terms and factual relations in specific domains, such as finance or law.
Example
When a customer service AI responds, 'The transaction fee for Product A is free,' the hallucination detection process compares the response against internal policy documents to verify if the statement is grounded. If there is insufficient grounding, it can hold the response or route it to a human agent for confirmation.