OncoAgent: A Dual‑Tier Multi‑Agent Framework for Privacy‑Preserving Oncology Clinical Decision Support
Implement the complexity router to route clinical queries to Tier 1 (9B) or Tier 2 (27B) models based on the weighted additive score.
Review the OncoAgent GitHub repo and assess its suitability for your oncology workflows.
Summary
OncoAgent is an open‑source clinical decision support system that combines a dual‑tier fine‑tuned LLM architecture with a multi‑agent LangGraph topology. The system routes queries through a complexity scorer to either a 9B speed‑optimised model (Tier 1) or a 27B deep‑reasoning model (Tier 2), both fine‑tuned via QLoRA on a corpus of 266,854 oncological cases using the Unsloth framework on AMD Instinct MI300X hardware. A four‑stage Corrective RAG pipeline over 70+ physician‑grade NCCN and ESMO guidelines achieves 100 % document grading success, while a three‑layer reflexion safety validator enforces a strict Zero‑PHI policy.
The architecture decomposes reasoning across eight LangGraph nodes, each with bounded, auditable functions. Human‑in‑the‑loop gates are mandatory for Tier 2 cases or low‑confidence outputs, and a fallback node returns a safe refusal when unrecoverable failures occur. Per‑patient memory isolation is achieved via unique thread IDs and LangGraph checkpointing. The entire stack runs natively on AMD MI300X using ROCm, enabling on‑premises deployment without cloud API dependency.
Key technical details include a weighted additive complexity model with specific weights for cancer type, stage, mutations, and prior treatment; a four‑stage retrieval pipeline (recall, distance gate, re‑ranking, context trimming); and embeddings from pritamdeka/S‑PubMedBert‑MS‑MARCO stored in a local ChromaDB index. Fine‑tuning completes in ~50 minutes, a 56× throughput acceleration over API‑based generation.
Key changes
- Dual‑tier architecture: 9B speed‑optimised model for Tier 1 and 27B deep‑reasoning model for Tier 2, both fine‑tuned via QLoRA
- Complexity router uses weighted additive model with specific weights for cancer type, stage, mutations, prior treatment; threshold S≥0.5 selects Tier 2
- Four‑stage Corrective RAG pipeline (recall, distance gate, re‑ranking, context trimming) achieves 100 % document grading success
- Three‑layer reflexion safety validator (formatting, safety, entailment) with deterministic critic node and max 2 retries
- Mandatory HITL gate for Tier 2 or low‑confidence outputs; fallback node returns safe refusal
- Per‑patient memory isolation via unique thread IDs and LangGraph checkpointing
- Knowledge base built from 70+ NCCN/ESMO guidelines, 266,854 cases, embeddings from pritamdeka/S‑PubMedBert‑MS‑MARCO, stored in local ChromaDB
- Full fine‑tuning on AMD Instinct MI300X (192 GB HBM3) completes in ~50 minutes, 56× throughput acceleration