OncoRAG

• 

  Interactive symptom intake chat interface grounded in WHO early-diagnosis guidance.

• 

  Sample doctor-ready PDF report with structured clinical intake summary and safety disclaimers.

• 

  One-click automated PDF summary generation for physician review.

Navigating early medical symptoms and complex oncology documentation can be overwhelming for patients. My goal was to build a safe, reliable AI intake assistant that helps patients structure their clinical history without hallucinating or providing unauthorized medical diagnoses.## Inspiration

OncoRAG is an evidence-grounded Retrieval-Augmented Generation (RAG) framework designed to streamline early medical intake:

• Guided Symptom Intake: Asks structured, empathetic follow-up questions based on WHO's Early Diagnosis Framework rather than jumping to conclusions.
• Hard Safety Guardrails: System prompts explicitly forbid diagnostic language and resist adversarial/prompt-injection attempts.
• Doctor-Ready PDF Export: Converts conversation logs into a clean clinical summary (Chief Complaint, HPI, Family History, Red Flags) for patients to bring to their appointment.

• Care Navigation: Provides a curated list of accredited oncology support centers and hotlines.## What it does

• Built with Python, LangChain, and OpenAI API for the core RAG framework.

• Utilized ChromaDB to store and query official WHO & ESMO early-diagnosis guidelines.

• Implemented strict system prompt architecture and guardrail checks to prevent diagnostic claims.

• Integrated automated PDF report generation (ReportLab) for structured clinical summaries.## How we built it

Ensuring zero diagnostic language while maintaining an empathetic tone was tough. Preventing prompt injections required rigorous adversarial testing to ensure the AI never breaches safety guardrails during role-play or emotional prompts.## Challenges we ran into

• Successfully grounding every response strictly in official WHO & ESMO medical standards.
• Achieving robust resilience against adversarial prompt-injection attempts.
• Creating a seamless flow from a live chat to a doctor-ready PDF summary.## Accomplishments that we're proud of

Building healthcare-adjacent AI requires prioritizing safety and evaluation over creative output. I learned how to evaluate LLM hallucination risks, handle edge cases, and implement strict RAG pipelines.## What we learned

• Expanding support for multilingual medical intake.
• Enhancing evaluation suites for edge-case adversarial prompts.
• Integrating direct exports to standard EHR (Electronic Health Record) formats.## What's next for OncoRAG

Built With

• chromadb
• langchain
• openai
• python
• rag