Medical AI that learns to ask better questions and verify its sources

Current medical question-answering systems use RAG to reduce hallucinations but rely on single-round retrieval that doesn't match clinical reasoning's iterative nature. SEMA-RAG addresses this by decomposing the task: an Interpreter Agent translates questions into clinically grounded queries, an Explorer Agent performs multi-round retrieval with feedback on whether evidence is sufficient, and an Arbiter Agent adjudicates conflicting sources before answering. Tested across five benchmarks and five LLM backbones, the approach improves accuracy by an average of 6.46 points over the strongest baselines.