AWS Machine Learning Blog

Category: Amazon Bedrock Agents

Flow diagram of custom hallucination detection and mitigation : The user's question is fed to a search engine (with optional LLM-based step to pre-process it to a good search query). The documents or snippets returned by the search engine, together with the user's question, are inserted into a prompt template - and an LLM generates a final answer based on the retrieved documents. The final answer can be evaluated against the reference answer from the dataset to get a custom hallucination score. Based on a pre-defined empirical threshold, a customer service agent is requested to join the conversation using SNS notification

Reducing hallucinations in large language models with custom intervention using Amazon Bedrock Agents

This post demonstrates how to use Amazon Bedrock Agents, Amazon Knowledge Bases, and the RAGAS evaluation metrics to build a custom hallucination detector and remediate it by using human-in-the-loop. The agentic workflow can be extended to custom use cases through different hallucination remediation techniques and offers the flexibility to detect and mitigate hallucinations using custom actions.