Clinical Drug Monograph RAG

Retrieval-Augmented Generation over a TNF-alpha inhibitor monograph using Gemini and ChromaDB

What It Does

A Retrieval-Augmented Generation (RAG) pipeline that lets you ask plain-language clinical questions about a drug monograph and get answers grounded strictly in the document. Built over a TNF-alpha inhibitor monograph as a concrete healthcare example.

Sample queries it answers correctly:

  • What is the initiation dose?
  • What are the contraindications?
  • What is the most severe potential side effect?
  • What doses are available?

Pipeline Architecture

RAG pipeline: PDF monograph loaded and chunked, embedded with Gemini using task-type-optimized vectors, stored in ChromaDB, and retrieved to ground Gemini 2.5 Flash answers to document content only.

Why It’s Interesting

Drug monographs are dense, technical, and long. Clinicians and pharmacists reference them frequently but navigating them is slow. A RAG layer turns a static PDF into an interactive Q&A surface — with answers that cite only what’s in the document, not hallucinated general knowledge.

The task_type distinction in the Gemini embedding API — retrieval_document for indexing, retrieval_query for querying — is a meaningful quality lever that most naive RAG implementations miss. The pipeline also uses a character budget on retrieved context to control generation cost, and a grounded prompt that restricts the model to retrieved chunks only.

Stack

  • Embeddings: Google Gemini gemini-embedding-001 (document + query optimized task types)
  • Vector store: ChromaDB (persistent)
  • Generation: Gemini 2.5 Flash
  • Document loading: LangChain + PyPDF + Google Cloud Storage
  • Chunking: RecursiveCharacterTextSplitter (1,000 chars, 100 overlap)

Notebook

The full implementation is available as a Colab-ready notebook:

View notebook on GitHub

Next Steps

  • Chunking by page first to avoid splitting mid-sentence across clinical sections
  • Biomedical embeddings (MedCPT, BiomedBERT) for better clinical terminology retrieval vs general-purpose Gemini embeddings
  • Multi-document extension across a full drug class (all TNF inhibitors) for comparative Q&A
  • Evaluation layer with a gold-standard Q&A set to measure retrieval precision and answer accuracy