pat-jj/s3

Trains search agents independently from generators using reinforcement learning, enabling modular RAG systems that work with any black-box LLM without retraining the language model itself. Integrates vLLM for efficient LLM serving, PySerini for retrieval, and FAISS for dense vector search, with support for multiple corpora (Wikipedia, MedCorp) and retrieval backends (e5 embeddings, BM25). Includes pre-configured baselines (RAG, DeepRetrieval, Search-R1, IRCoT) and provides checkpoint initialization via naive RAG caching to minimize training data requirements.

820 stars.