Bayesian Uncertainty Improves Agentic RAG Pipeline Trustworthiness

Deploying AI agents in real-world scenarios, especially those involving multi-stage reasoning like Agentic RAG pipelines, requires robust mechanisms to identify when the system might fail. This research proposes an uncertainty-aware framework that integrates Bayesian Networks to track and propagate uncertainty signals across different stages of an Agentic RAG system. These signals originate from components like the planner, evaluator, and generator, based on factors such as semantic divergence and self-evaluation.By using a Bayesian Network, the framework can estimate overall system uncertainty and pinpoint specific stages where failures are likely to occur. Evaluated on multi-hop question-answering datasets, the approach showed particular effectiveness in scenarios where uncertainty accumulates across reasoning steps. This preliminary study suggests a promising method for monitoring and improving the reliability of complex AI agent workflows.