A framework for experimenting with and optimizing Retrieval-Augmented Generation (RAG) pipeline architectures.

This package provides a framework for testing different RAG pipeline configurations and measuring their performance. It supports pluggable components for document ingestion, chunking strategies, and retrieval methods.

• Modular Architecture: Pluggable components across the ingestion and retrieval stack with sample wrappers over top open-source component offerings.
• Experiment Pipeline: Generate and test all combinations of component variants
• LangChain Integration: RAGArtifact extends BaseRetriever for seamless use with LangChain chains

pip install rag_select

To set up an experiment with RAG-select, follow these steps:

1. Prepare Your Dataset and Documents

   • dataset: This should be your evaluation set (e.g., queries and ground-truths).
   • documents: The corpus to be indexed and retrieved from.

2. Define Component Variants

   • Specify different variants for each pipeline stage. Import and instantiate implementations as needed.

   Example:

   from rag_select.parameter_impls.chunking_impls import SlidingWindowChunking
   from rag_select.parameter_impls.embedding_impls import HuggingFaceEmbedding
   from rag_select.parameter_impls.retriever_impls import SimpleRetriever
   
   chunking_variants = [
       SlidingWindowChunking(chunk_size=256, chunk_overlap=20),
       SlidingWindowChunking(chunk_size=512, chunk_overlap=50),
   ]
   embedding_variants = [
       HuggingFaceEmbedding(model_name="sentence-transformers/all-MiniLM-L6-v2"),
   ]
   retriever_variants = [
       SimpleRetriever(top_k=3),
       SimpleRetriever(top_k=5),
   ]

3. Run the Experiment

   • Instantiate the RAGExperiment and call .run() to evaluate all pipeline combinations.
   • Review, compare, or rank the results.

   Example:

   from rag_select.experiment.rag_experiment import RAGExperiment
   
   experiment = RAGExperiment(
       dataset=eval_dataset,
       documents=documents,
       search_space={
           "chunking": chunking_variants,
           "embedding": embedding_variants,
           "retriever": retriever_variants,
       },
       metrics=["precision@3", "precision@5", "recall@5", "mrr"],
   )
   
   results = experiment.run()
   
   # Rank pipelines by a metric
   ranked = results.rank(by="precision@5")
   
   # Get the best pipeline config
   best = results.get_best_pipeline()

4. Extending the Search Space

   • To experiment with ingestion or storage variants, include them as keys in search_space and provide corresponding implementations.

MIT