PyTerrier RAG

pyterrier-rag is an extension for PyTerrier that makes it easier to produce retrieval augmented generation pipelines. PyTerrier-RAG supports:

1. Easy access to common QA datasets

2. Pre-built indices for common corpora

3. Popular reader models, such as Fusion-in-Decoder, LLama

4. Evaluation measures

As well as access to all of the retrievers (sparse, learned sparse, and dense) and rerankers (from MonoT5 to RankGPT) accessible through the wider PyTerrier ecosystem.

• RAG Datamodel
• RAG Measures
• LLM Backends
• Prompt Construction
• Readers

Example Notebooks

Try out the following example notebooks to get started with PyTerrier RAG:

• Sparse Retrieval on Natural Questions with FiD and FlanT5 readers: sparse_retrieval_FiD_FlanT5.ipynb

• Sparse Retrieval on Natural Questions with Mistral: sparse_retrieval_Mistral.ipynb

• E5 Dense Retrieval with FiD on Natural Questions: dense_e5_retrieval_FiD.ipynb

• KG-augmented RAG with TRACE on HotpotQA: trace.ipynb

• Agentic RAG: R1-Searcher r1searcher.ipynb

• Agentic RAG: Search-R1 search-r1.ipynb

• Agentic RAG: Search-O1 search-o1.ipynb

Credits

• Craig Macdonald, University of Glasgow

• Jinyuan Fang, University of Glasgow

• Andrew Parry, University of Glasgow

• Zaiqiao Meng, University of Glasgow

• Sean MacAvaney, University of Glasgow