OncoAgent: Open-source multi-agent framework for oncology decision support launches with privacy-preserving architecture

OncoAgent is an open-source clinical decision support system designed specifically for oncology, built as a directed graph topology using LangGraph with eight specialized nodes. The architecture decomposes clinical reasoning into bounded, auditable functions rather than relying on a single monolithic model, addressing the core challenge of context saturation in complex multi-comorbidity cases.

The system routes incoming oncology queries through a complexity scorer that assigns cases to one of two fine-tuned models: a 9-billion-parameter speed-optimized tier for straightforward triage, or a 27-billion-parameter reasoning-focused tier for Stage IV cancers, rare diagnoses, or multiple genetic mutations. The routing decision combines weighted factors including cancer type, disease stage, identified mutations, and prior treatment history, with manual clinician override available through the user interface.

All model outputs are grounded in a curated knowledge base of 70+ clinical guidelines from the National Comprehensive Cancer Network (NCCN) and European Society for Medical Oncology (ESMO) through a four-stage retrieval-augmented generation pipeline. The system automatically grades retrieved documents for clinical relevance before forwarding to the specialist node; documents that fail relevance classification trigger query reformulation to eliminate the primary hallucination source in RAG systems.

Safety enforcement operates across three layers: a reflexion loop that self-corrects generation via feedback-augmented retry (maximum two iterations), a critic node that validates outputs against safety criteria, and a human-in-the-loop gate that mandates clinician review for high-complexity cases or low-confidence outputs. The entire system enforces a strict zero-PHI policy, preserving complete audit trails through immutable state representation.

Both models underwent dual-tier QLoRA fine-tuning on a corpus of 266,854 real and synthetically generated oncological cases using the Unsloth framework optimized for AMD Instinct MI300X hardware (192 GB HBM3 memory). Sequence packing enabled full-dataset fine-tuning to complete in approximately 50 minutes, representing a 56-fold acceleration over prior API-based generation workflows. Post-optimization, corrective RAG document grading achieved 100% success rate with a mean RAG confidence score of 2.3 or higher.

The system runs entirely on-premises using AMD ROCm and open-source frameworks, eliminating dependency on proprietary cloud APIs and eliminating risk of patient data exfiltration—a requirement in privacy-sensitive hospital environments. The complete codebase is published as open source.