LLM-based document reranking approach that fine-tunes decoder-only models like LLaMA to calculate query-document relevance scores. Uses generative capabilities of large language models to improve retrieval ranking in search and RAG systems.
RankGPT is a reranking approach that leverages decoder-only large language models (LLMs) such as LLaMA to improve document ranking in retrieval systems. It represents one of the supervised LLM reranker methods, where pre-trained LLMs are fine-tuned on ranking-specific datasets like MS MARCO to develop ranking awareness that is absent during standard pre-training.
RankGPT formulates document reranking by fine-tuning decoder-only language models. Different from encoder-decoder approaches like RankT5 which generate classification tokens, RankGPT uses decoder-only architectures for relevance calculation. It employs prompting strategies to have LLMs improve document reranking autonomously.
RankGPT can utilize different prompting strategies:
Zero-shot LLM-based rerankers like RankGPT exhibit competitive effectiveness, with some matching the performance of GPT-3.5 Turbo on various datasets. However, inefficiency and high costs currently limit their practical deployment in production retrieval systems compared to cross-encoder approaches.
Loading more......