UPR Reranker

Overview

The UPRReranker class is a reranker based on paper called "Improving Passage Retrieval with Zero-shot Question Generation". It uses a language model to generate a question based on the passage and reranks the passages by the likelihood of the question.

It can enhance accuracy because it calculates likelihood original passages and related passages (generated) with user's question. Also, it is cost-effective than LLM Reranker because it doesn't generate whole passages.

Usage

To initialize an instance of UPRReranker:

from RAGchain.reranker import UPRReranker

# if use gpu
reranker = UPRReranker(use_gpu=True)

To rank given set of passages:

query ="What causes global warming?"
passages =[...list_of_passages...] # Assume we have list_of_passages retrieved earlier

reranked_passages= reranker.rerank(query=query, passages=passages)
print(reranked_passages)