I need a retrieval-augmented large-language-model that can (1) suggest differential diagnoses, (2) explain conditions and treatments in plain language for professionals and patients alike, and (3) surface novel insights to accelerate ongoing research projects. The scope is intentionally broad—chronic diseases, infectious diseases, and mental-health conditions must all be handled with equal rigor. The system will ingest three primary data streams: peer-reviewed medical literature, de-identified patient health records, and curated clinical-trial datasets. I already have secure access paths for each source; what I lack is the unified pipeline that cleans, embeds, and indexes the content so the model can ground every answer in verifiable evidence. Python, LangChain (or comparable orchestration), Hugging Face transformers, and a vector store such as FAISS or Pinecone feel like the natural toolkit here, but I am open to persuasive alternatives if they improve latency or compliance. HIPAA-level security and full audit trails are non-negotiable. Deliverables • Data-ingestion and cleansing pipeline connected to all three data sources • Vector index with citations back to the original documents or EHR entries • Fine-tuned or custom-trained LLM with RAG architecture • API endpoints (REST or gRPC) plus a lightweight web demo for clinical reviewers • Evaluation report covering diagnostic accuracy, factual consistency, and safety filters Acceptance criteria: every response cites its sources, PII never leaks, and benchmark tests meet or exceed baseline scores we will define together before training begins.