How novasearch-team/rag-retrieval works
Project type and purpose
• RAG-Retrieval is a research-oriented full-stack toolkit and library for “Retrieval-Augmented Generation” (RAG) systems.
• It provides end-to-end support for:
a. Training RAG retrieval components: embedding models, late-interaction (ColBERT) models, and reranker models (both BERT and LLM-based).
b. Inference: a lightweight Python library (rag-retrieval) offering a unified interface to serve any reranker model.
c. Distillation: compressing large embedding or reranker models into smaller, production-friendly ones.
File structure
• Top-level:
• rag_retrieval/ – core Python package, split into:
– rag_retrieval/train/ – submodules for each training pipeline (colbert, embedding, reranker).
– rag_retrieval/infer/ – inference API, reranker implementations.
– rag_retrieval/reranker.py – high-level entrypoint.
• config/ – DeepSpeed and FSDP distributed-training configurations.
• example_data/, examples/ – sample inputs, end-to-end experiment scripts.
• Config files and scripts (*.sh) reveal distributed multi-GPU support via DeepSpeed/FSDP.
• tests/ validate each reranker variant.
README for additional insights
• README describes three pillars (training/inference/distillation), lists supported model families (bert, LLM, colbert), and points to per-pipeline READMEs.
• Emphasizes modularity: new rerankers plug into BaseReranker.
Final advice
• Be detailed: show all three training modules and the unified inference module.
• Distributed training configurations separately.
• Modularity and plugin points.