Document-Grounded Medical Question Answering From Uploaded PDFs

Clinicians, medical students, and healthcare administrators work from dense PDFs: clinical guidelines, drug references, research papers, and policy documents often run to hundreds of pages. Finding the answer to a specific question means manual searching through material where precision matters and paraphrase can be dangerous. A question answering assistant is only acceptable in this setting if every answer is demonstrably grounded in the uploaded documents themselves. The intern builds exactly that assistant. Users upload medical PDFs through a Python Flask application; custom data loaders built with LlamaIndex parse the documents, preserving structure such as sections and tables, and store embedded chunks in a vector store. When a question arrives, a LangChain retrieval-augmented generation pipeline fetches the most relevant passages and OpenAI models compose an answer constrained to that retrieved context, with source citations pointing back to the document and page so users can verify every claim. The intern engineers the chunking strategy, the retrieval tuning, and the guardrails that make the system refuse to answer beyond its sources. The project demonstrates the most employable pattern in applied generative AI, high-stakes RAG: custom ingestion, vector store design, grounded generation, and the safety-first engineering judgment that healthcare software demands.