Hybrid Agentic-RAG Framework for Healthcare Decision Support

Inspiration

Healthcare AI tools often rely on either predictive models or retrieval-based language models, but rarely combine both in a principled manner. Predictive systems lack transparent explanations, while RAG-only systems struggle to reason over patient-specific symptom patterns and uncertainty. This gap motivated us to design a unified framework that can jointly reason over structured symptom data and authoritative medical knowledge, while explicitly accounting for confidence and uncertainty in medical decision support.

What it does

The Hybrid Agentic–RAG Framework offers healthcare decision support by integrating symptom-based disease prediction with evidence-based medical explanations. It accepts natural language queries, determines whether the query is symptom-driven or informational, extracts structured symptoms when needed, predicts probable conditions with confidence scores, and generates grounded explanations using retrieved medical knowledge. The system adapts its reasoning strategy based on confidence, producing focused explanations for high-confidence cases and safer, general guidance when uncertainty is high

How we built it

We built the system using a modular agentic architecture implemented with LangChain and LangGraph. A trained machine learning prediction agent infers diseases from structured symptom inputs under realistic noise conditions, while two ChromaDB vector stores support retrieval from structured symptom knowledge and an unstructured medical encyclopedia. An orchestrator coordinates intent routing, confidence-aware decision logic, and prompt construction. The backend is implemented using FastAPI, and a React-based frontend provides a multi-session chat interface with persistent conversational state.

Challenges we ran into

A major challenge was avoiding overconfident predictions when user-provided symptoms were incomplete or ambiguous. Initial experiments with low symptom noise yielded unrealistically high accuracy, suggesting overfitting to complete symptom patterns. Another challenge was balancing retrieval depth and response quality, as smaller chunk sizes and low retrieval limits resulted in fragmented or insufficient medical context. Designing a confidence-aware orchestration strategy that remained informative while prioritizing safety was also a nontrivial task.

Accomplishments that we're proud of

We successfully built a confidence-aware agentic system that integrates structured machine learning inference with retrieval-augmented generation in a transparent and modular way. The prediction agent achieved a mean accuracy of 94.84 percent under realistic noise using four-fold cross-validation, while retrieval evaluation showed consistent improvements in Recall@K as more context was retrieved. Most importantly, the system demonstrates safe fallback behavior under uncertainty, which is critical for healthcare-facing applications.

What we learned

This project showed that predictive models and RAG are complementary rather than competing approaches. Explicit agentic control enables clearer reasoning flow, better interpretability, and safer behavior than monolithic pipelines. We also learned that confidence signals are essential in healthcare decision support, not just for accuracy reporting but for controlling how explanations are generated and how strongly conclusions are stated.

What's next for the Hybrid Agentic-RAG Framework for Healthcare Decision Support

Future work includes expanding the structured dataset to support a broader range of conditions, improving symptom extraction robustness for vague user inputs, and integrating additional medical knowledge sources. We also plan to explore richer evaluation with clinician-in-the-loop feedback and extend the framework to other high-stakes domains where confidence-aware reasoning is required over structured and unstructured data.

Built With

• chromadb
• fastapi
• hugging-face-models
• javascript
• langchain
• langgraph
• llm
• openai
• python
• react
• vector-databases