Core Concepts Overview

Understanding RAG systems for medical question-answering in obstetrics

What is Retrieval-Augmented Generation?

Retrieval-Augmented Generation (RAG) is a powerful technique that enhances Large Language Models (LLMs) by combining them with external knowledge retrieval. Instead of relying solely on a model’s pre-trained knowledge, RAG systems:

1. Retrieve relevant information from a knowledge base
2. Augment the model’s context with this information
3. Generate accurate, grounded responses

Why RAG for Medical Q&A?

Medical information requires high accuracy and up-to-date knowledge. RAG systems ground LLM responses in verified medical documentation, reducing hallucinations and ensuring answers are based on authoritative sources.

The Obstetrics RAG Benchmark

This project systematically evaluates different RAG architectures for answering questions about pregnancy, prenatal care, and childbirth. The benchmark uses real medical guidance documents to create a comprehensive test of RAG effectiveness in the healthcare domain.

Research Objectives

Architecture Comparison

Evaluate 6 distinct RAG retrieval strategies from simple semantic search to advanced query reformulation

Model Performance

Assess how different LLMs (GPT-4o default, plus GPT-5, MediPhi, MedGemma) perform with each RAG architecture

Retrieval Quality

Measure precision and recall of different retrieval strategies for medical content

Best Practices

Identify optimal configurations for medical question-answering systems

System Architecture

The Obstetrics RAG Benchmark implements a complete evaluation pipeline:

Key Components

Data Pipeline

Raw medical documents are processed, chunked, and embedded into a ChromaDB vector store. This creates a searchable knowledge base optimized for semantic retrieval.

Processing Steps:

• Document extraction from PDFs
• Text chunking with overlap
• Embedding generation (OpenAI text-embedding-3-small)
• Vector store indexing

RAG Architectures

Six different retrieval strategies are implemented, each representing a distinct approach to finding relevant medical information:

• Simple Semantic Search (baseline)
• Hybrid Search (BM25 + Semantic)
• Hybrid with RRF (Reciprocal Rank Fusion)
• HyDE (Hypothetical Document Embeddings)
• Query Rewriter (Multi-Query)
• PageIndex (External API)

Language Models

The system supports evaluation across multiple language models:

Default models (used by RAG implementations):

• GPT-4o: Default high-performance model (temperature=0)
• GPT-3.5-turbo: Used for HyDE hypothetical document generation

Multi-model evaluation registry (via MODELS_REGISTRY):

• gpt-5, gpt-5.2: Next-generation OpenAI models
• microsoft/MediPhi-Instruct (mediphi): Medical-specialized HuggingFace model
• google/medgemma-1.5-4b-it (medgemma): Medical-focused compact model

Evaluation Framework

RAGAS (Retrieval-Augmented Generation Assessment) provides automated, LLM-based evaluation metrics that assess both retrieval quality and answer quality without manual annotation.

How RAG Components Work Together

A typical RAG query flows through multiple stages:

1. Query Processing

The user’s question is analyzed and potentially transformed depending on the RAG architecture:

# Example: Simple semantic search
query = "¿Cuál es la cantidad ideal de controles prenatales?"

# Example: Query rewriting
queries = [
    "¿Cuál es la cantidad ideal de controles prenatales?",
    "¿Cuántas consultas prenatales se recomiendan?",
    "Número óptimo de visitas durante el embarazo"
]

2. Document Retrieval

Relevant documents are retrieved from the vector store:

# Semantic similarity search
retrieved_docs = vectorstore.similarity_search(query, k=5)

# Each document contains:
# - page_content: The actual text
# - metadata: Source, page number, chunk information

3. Context Construction

Retrieved documents are formatted as context for the LLM:

--- Document 1 ---
Source: Guía Práctica Clínica, Page: 15
Content: Se recomienda un programa de diez citas...

--- Document 2 ---
Source: Guía Práctica Clínica, Page: 16
Content: Para una mujer multípara...

4. Answer Generation

The LLM generates a response grounded in the retrieved context:

response = llm.invoke(
    f"Context: {formatted_context}\n\nQuestion: {query}\n\nAnswer:"
)

5. Evaluation

The answer is evaluated against ground truth using RAGAS metrics:

• Faithfulness: Is the answer grounded in the context?
• Answer Relevancy: Does it address the question?
• Context Precision: Were relevant documents retrieved?
• Context Recall: Was all necessary information retrieved?

Research Dataset

The benchmark uses a curated dataset of 10 questions about pregnancy and prenatal care, each with ground truth answers derived from clinical practice guidelines:

Example Question

Question: ¿Cuál es la cantidad ideal de controles prenatales?

Ground Truth: Se recomienda un programa de diez citas. Para una mujer multípara con un embarazo de curso normal se recomienda un programa de siete citas.

Evaluation: The system retrieves relevant guidelines and generates an answer, which is then scored on faithfulness, relevancy, precision, and recall.

Key Research Questions

This benchmark addresses critical questions for medical RAG systems:

1. Which retrieval strategy produces the most accurate medical answers?
   • Does semantic search alone suffice, or do hybrid approaches improve results?
2. How much does model choice matter?
   • How do next-gen models (GPT-5) compare to GPT-4o for medical RAG?
   • Can specialized medical models (MediPhi, MedGemma) outperform general-purpose LLMs?
3. What are the cost-performance tradeoffs?
   • Can cheaper models achieve similar quality with better retrieval?
4. Do advanced techniques justify their complexity?
   • Are HyDE and query rewriting worth the additional LLM calls?

Next Steps

RAG Architectures

Explore the 6 different retrieval strategies in detail

Evaluation Framework

Learn how RAGAS measures RAG system quality

Data Pipeline

Understand document processing and vector store creation

Getting Started

Run your first evaluation