ClinicBot: A Guideline-Grounded Clinical Chatbot with Prioritized Evidence RAG and Verifiable Citations
Fights hallucination with guideline-grounded evidence prioritization for diabetes care.
ClinicBot, developed by Navapat Nananukul and Mayank Kejriwal, tackles hallucination in medical AI by grounding answers in official guidelines. Unlike standard RAG systems that treat all retrieved evidence equally, it employs three key advances: (1) structured extraction of clinical guidelines into semantic units (recommendations, tables, definitions, narrative) with provenance, (2) an evidence prioritization mechanism that ranks content by clinical importance rather than textual similarity, and (3) a web interface offering concise, actionable answers with verifiable citations.
The system demonstrated on diabetes questions from real patients, using the American Diabetes Association Standards of Care (2025). It also includes a diabetes risk assessment tool. The multi-agent architecture processes complex guidelines at scale, ensuring outputs align with clinical practice. By combining semantic extraction and hierarchical ranking, ClinicBot aims to make AI trustworthy in high-stakes medical settings, providing doctors and patients with reliable, cited information.
- Structured extraction parses ADA 2025 guidelines into recommendations, tables, definitions, and narrative with explicit provenance.
- Evidence prioritization ranks content by clinical significance, not just textual similarity, reducing noisy context.
- Multi-agent setup scales to process complex guidelines and delivers verifiable citations for every answer.
Why It Matters
ClinicBot addresses medical AI hallucination by grounding answers in prioritized, cited guidelines, boosting trust for clinicians.