RankT5

Overview

RankT5 is a supervised LLM reranker that employs a T5 encoder-decoder architecture for document reranking. It formulates the ranking problem as a text generation task, outputting classification tokens to indicate relevance.

Architecture

• Model Type: Encoder-decoder (T5) backbone
• Approach: Pointwise or pairwise ranking formulation as a generation task
• Output: Generates classification tokens (e.g., "relevant" or "irrelevant") for query-document pairs

Training

• Supervised fine-tuning on passage ranking datasets such as MS MARCO
• Adjusts pre-trained T5 parameters for improved relevance measurement
• Addresses the lack of ranking awareness inherent in pre-trained language models

Use Cases

• Search result reranking
• RAG pipeline document prioritization
• Semantic relevance scoring for query-passage pairs

Performance

• Competitive with other LLM-based rerankers on standard benchmarks
• More efficient than listwise LLM reranking methods that require sliding window strategies
• Open-source and can be self-hosted, avoiding API costs