Skip to main content
This guide covers how to extend the Obstetrics RAG Benchmark project with new data sources, retrieval parameters, and research contributions.

​
Extension Points

The benchmark provides several extension points for researchers:
  • RAG Architectures: Add novel retrieval strategies (~/workspace/source/src/rag/)
  • Model Integration: Test new LLMs and SLMs (~/workspace/source/src/common/model_provider.py)
  • Evaluation Metrics: Extend RAGAS evaluation (~/workspace/source/src/evaluation/ragas_evaluator.py)
  • Data Sources: Integrate new medical corpora (~/workspace/source/data/)

​
Adding New Data Sources

The benchmark uses a medical corpus on pregnancy and childbirth. To add new data sources:
1

Prepare Document Chunks

Create a JSON file in data/chunks/ with the following structure:
[
  {
    "content": "Your medical text content here...",
    "source": "document_name.pdf",
    "page_number": 1,
    "chunk_id": "chunk_001"
  }
]
Each chunk should contain:
  • content: The text content of the chunk
  • source: Original document filename
  • page_number: Page number in source document
  • chunk_id: Unique identifier for the chunk
2

Generate Embeddings

Run the embedding generation script to create vector representations:
python scripts/create_embeddings.py
This will:
  • Load chunks from data/chunks/chunks_final.json
  • Generate embeddings using OpenAI’s text-embedding-3-small
  • Store them in ChromaDB at data/embeddings/chroma_db/
3

Update Collection Name

If using a different medical domain, update the collection name in RAG implementations:
# In src/rag/simple.py, hybrid.py, etc.
collection_name = "your_collection_name"

vectorstore = Chroma(
    persist_directory=str(chroma_db_dir),
    embedding_function=embeddings,
    collection_name=collection_name,
)
4

Update Ground Truth Questions

Modify the evaluation dataset in src/evaluation/ragas_evaluator.py:49-90:
DATA_GT = [
    {
        "question": "Your domain-specific question?",
        "ground_truth": "Expected answer based on your corpus."
    },
    # Add more question-answer pairs...
]

​
Modifying Retrieval Parameters

Each RAG implementation allows retrieval parameter tuning:

​
Adjusting Number of Retrieved Documents

# In src/rag/simple.py:54
retriever = vectorstore.as_retriever(search_kwargs={"k": 5})  # Change k value

# In src/rag/hybrid.py:66,74
bm25_retriever.k = 5  # Adjust BM25 retrieval count
semantic_retriever = vectorstore.as_retriever(search_kwargs={"k": 5})

​
Tuning Hybrid Retrieval Weights

# In src/rag/hybrid.py:77-82
ensemble_weight_bm25 = 0.5      # Lexical weight
ensemble_weight_semantic = 0.5   # Semantic weight

ensemble_retriever = EnsembleRetriever(
    retrievers=[bm25_retriever, semantic_retriever],
    weights=[ensemble_weight_bm25, ensemble_weight_semantic]
)
Best Practice: Document parameter changes in evaluation metadata to track their impact on performance.

​
Modifying Temperature for Generation

# For creative generation (HyDE hypothetical documents)
llm_hyde = ChatOpenAI(model_name="gpt-3.5-turbo", temperature=0.7)

# For deterministic answers
llm_answer = ChatOpenAI(model_name="gpt-4o", temperature=0)

​
Experimentation Best Practices

​
1. Version Control Your Experiments

Create separate branches for experimental changes: 
git checkout -b experiment/new-chunking-strategy

​
2. Track Configuration Changes

Document experiments in your evaluation metadata: 
metadata = {
    "experiment_name": "increased_context_window",
    "retrieval_k": 10,
    "changes": "Doubled retrieved documents from 5 to 10",
    "hypothesis": "More context improves faithfulness scores"
}

​
3. Run Comparative Evaluations

Always compare against baseline: 
# Baseline
python scripts/run_evaluation.py simple

# Your modification
python scripts/run_evaluation.py simple

​
4. Analyze Results Systematically

Compare metrics across configurations:
  • Faithfulness: Did more context reduce hallucinations?
  • Answer Relevancy: Are answers still focused on the question?
  • Context Precision: Is the retrieved context more relevant?
  • Context Recall: Are we capturing all necessary information?

​
Contributing to the Project

We welcome research contributions that advance RAG techniques for medical Q&A:

​
Contribution Areas

Novel RAG Architectures

Implement and evaluate new retrieval strategies

Model Integration

Add domain-specialized medical language models

Evaluation Extensions

Propose additional metrics or analysis methods

Results & Analysis

Contribute comparative insights and visualizations

​
Contribution Workflow

1

Fork the Repository

git clone https://github.com/NicolasHoyosDevs/RAG-Benchmark.git
cd RAG-Benchmark
2

Create Feature Branch

Use descriptive names for your branch:
git checkout -b feature/semantic-reranking
3

Implement Your Changes

Follow the existing code structure and patterns. Add comprehensive docstrings.
4

Run Evaluations

Evaluate your changes across multiple models:
python scripts/run_evaluation.py multi-model your-rag-type
5

Document Your Methodology

Include:
  • Hypothesis and motivation
  • Implementation details
  • Experimental setup
  • Results summary and analysis
6

Submit Pull Request

Create a PR with:
  • Clear description of changes
  • Evaluation results comparison
  • Analysis of improvements/trade-offs

​
Research Guidelines

​
Reproducibility

  • Fix Random Seeds: Set random seeds for reproducible results
  • Document Dependencies: Update requirements.txt with new packages
  • Save Configurations: Store all hyperparameters in configuration files

​
Statistical Rigor

  • Run multiple trials to account for variance
  • Report mean and standard deviation for metrics
  • Use appropriate statistical tests for comparisons

​
Ethical Considerations

When extending to new medical domains:
  • Ensure data sources are properly licensed
  • Validate medical accuracy with domain experts
  • Include appropriate disclaimers about clinical use
  • Respect patient privacy in any real-world data

​
Next Steps

Adding RAG Architectures

Learn how to implement new RAG strategies

Integrating Models

Add new LLMs to the model registry

Customizing Metrics

Extend evaluation with custom metrics

API Reference

Explore the complete API documentation

Build docs developers (and LLMs) love

Get started for freeTalk to us