OncoAgent: A Multi-Agent Framework for Oncology

OncoAgent is an open-source framework designed to support complex clinical decision-making in the field of oncology. It combines a LangGraph-based multi-agent architecture with a Corrective RAG (CRAG) pipeline to provide answers grounded in expert guidelines such as NCCN and ESMO.

Key features are as follows.

• Dual-Tier LLM Architecture: Depending on query complexity, it routes between a 9B model (optimized for speed) and a 27B model (for deep reasoning) to maximize efficiency.
• Data Privacy and Security: It applies a strict Zero-PHI policy and a 3-stage Reflexion safety verifier, and supports on-premises deployment to prevent patient data from leaking externally.
• High-Performance Training and Hardware: It performed QLoRA fine-tuning on over 260,000 cases using AMD Instinct MI300X hardware and the Unsloth framework.
• Reliable Generation: A 4-stage RAG pipeline evaluates document relevance and minimizes hallucinations, increasing its usability in clinical settings.