The Medical Hallucination Problem
When deploying LLMs in clinical contexts, accuracy is not optional. Standard language models generate plausible-sounding text, but they can occasionally misstate drug quantities or pathway side effects—an unacceptable outcome for patient-facing software.
Grounding with Google OKF
To enforce factual alignment, our models do not rely solely on internal weights. Instead, they reference structured semantic guidelines mapped from the Google Open Knowledge Foundation (OKF) schemas.
This workflow guarantees:
- **Entity Resolution**: Resolving synonyms (e.g. mapping "Aspirin" and "Acetylsalicylic acid" to a single RxNorm ID).
- **Relationship Validation**: Verifying compound-disease connections using graph verification algorithms.
- **Traceable Citations**: Restricting generator outputs to verified scientific literature blocks.
System Architecture
Our engine utilizes a two-step RAG verification loop:
1. Retrieval: Extract vector indices alongside knowledge graph pathways.
2. Validation: Check generated responses against active OKF relationship rules before rendering.