RAG-Retrieval

Why we included this project

Teams fine-tuning a retrieval pipeline on their own domain data will find a practical training kit here, not just another inference wrapper. The project covers the three model families that show up in most RAG setups: dense embedding models, late-interaction ColBERT-style encoders, and cross-encoder or LLM-based rerankers, all trainable against open checkpoints like BGE, GTE, and BCE. It also handles distillation, so you can shrink a large LLM reranker or embedding model down to a 0.5B or BERT-base size that costs less to host. On the inference side, a lightweight package offers one ranking interface for calling different rerankers, with long-document handling that helps when your corpus exceeds normal context limits. Researchers comparing retrieval approaches will also appreciate the positional-bias benchmarks and MTEB reranking results as a reference point for what the code can produce.