PriHA: A RAG-Enhanced LLM Framework for Primary Healthcare Assistant in Hong Kong
New framework tackles fragmented medical guidelines with a novel 'Dual RAG' architecture for reliable, localized answers.
A research team from Hong Kong has developed PriHA (Primary Healthcare Assistant), a new AI framework designed to solve a critical local problem: fragmented and inaccessible official healthcare information. To combat rising public health costs, the Hong Kong government encourages self-care, but guidelines are scattered across departments in various formats. General-purpose LLMs like ChatGPT often fail here, hallucinating or providing generic advice. PriHA addresses this by building a specialized system around a novel Dual Retrieval Augmented Generation (DRAG) architecture.
PriHA's core innovation is its tri-stage pipeline. First, a query optimizer breaks down a user's health question into intent-oriented sub-queries. The DRAG system then performs mixed-source retrieval, pulling relevant context from a curated database of localized Hong Kong medical documents. Finally, it reorganizes this context for generation, ensuring the AI's answers are grounded in official sources. Comprehensive experiments show PriHA outperforms both its own ablations and baseline models in accuracy and response clarity.
The framework provides more than just accurate answers; it offers traceability. Users and healthcare providers can see the source documents used to generate a response, increasing trust in a high-stakes domain. The researchers position PriHA as a blueprint for other localized, high-risk AI applications where reliability is non-negotiable. By successfully anchoring an LLM to verified, region-specific data, the project demonstrates a practical path forward for deploying AI in sensitive public service sectors beyond healthcare.
- Uses a novel Dual RAG (DRAG) architecture for mixed-source retrieval from official Hong Kong medical documents.
- Outperforms general models like ChatGPT in accuracy and clarity for localized healthcare queries, as validated in experiments.
- Provides a traceable dialogue framework, allowing users to see the source of generated medical advice.
Why It Matters
It demonstrates a reliable blueprint for deploying AI in high-risk, localized public services where accuracy is critical.