JD Health Replaces RAG with Deterministic Engineering Stack for Medical AI

Problem: Probabilistic AI Falls Short in High-Stakes Healthcare

Medical AI is shifting from answering questions to completing tasks, but in healthcare every output must be deterministically correct. General-purpose LLMs deployed directly suffer from three issues: retrieval results are probabilistically uncontrollable, models excel at answering but not at executing, and generic engineering frameworks cannot meet domain requirements — errors go undetected and unaccounted for.

The bottleneck is not model capability but the lack of determinism in the engineering stack from model to knowledge to engineering to delivery.

Solution: Four-Layer Engineering Stack Reconstruction

JD Health redesigned the entire stack across four layers:

Model Layer: Execution-Grade Medical Code LLM

Upgraded from a reasoning model to an execution-grade medical code LLM. Code becomes the carrier for task execution, producing outputs that are verifiable and traceable.

Knowledge Layer: Deterministic Agentic Search and Structured Knowledge Base (LLMwiki)

Removed probabilistic RAG. Built a deterministic Agentic Search and a structured knowledge base called LLMwiki. Verification follows logical paths, making sources traceable and the verification process auditable.

Harness Layer: Directed Harness (MedWork)

Developed a directed harness named MedWork. It organizes tasks and deliverables, constrains tool calls, and validates key medical logic.

Delivery Layer: Closed-Loop Evaluation, Execution Records, Error Recovery, and Human Review

Established a closed loop with evaluation, execution logging, error recovery, and human review. For long-running tasks, checkpoints and recovery mechanisms are set; critical results undergo professional review.

Deployment and Results

This engineering stack now powers MedWork across multiple clinical tasks: evidence retrieval, medical record summarization, outpatient assistance, recording-assisted diagnosis, AI-MDT (multidisciplinary team), clinical nutrition, medication assistance, imaging, and research.

Four self-developed components — Jingyi Qianxun 3.0, MedCode, LLMwiki, and MedWork — are fully integrated, pushing AI delivery from "high probability of being correct" to "deterministically correct."