AI
Single Model Unites Embedding and Reranking for Faster, Better Search Results
Researchers introduce E2Rank, a framework that combines text embedding and listwise reranking in one model to match LLM performance while cutting latency dramatically.
E2Rank offers a new approach to search by merging two traditionally separate stages—initial retrieval and reranking—into a single unified model, bridging the gap between speed and ranking accuracy.
The Challenge#
Document retrieval systems face a fundamental trade-off. Text embedding models are fast and efficient at finding relevant documents from large collections, but they struggle to rank results with the precision of LLM-based rerankers. Those large language models can capture complex interactions between queries and documents, yet they require expensive autoregressive decoding for each ranking decision, making them prohibitively slow.
How E2Rank Works#
The technique extends a standard text embedding model through continued training on a listwise ranking objective. The innovation lies in treating a reranking prompt—built from the query and its top K candidates—as pseudo-relevance feedback (PRF). This approach allows the model to rerank documents by computing cosine similarity against precomputed embeddings, eliminating the need for slow token-by-token generation.
Performance and Efficiency#
Testing across multiple benchmarks shows strong results:
- Achieves state-of-the-art performance on BEIR, a standard retrieval benchmark
- Delivers competitive outcomes on BRIGHT, a reasoning-heavy benchmark
- Produces significantly lower latency than comparable LLM-based rerankers
- Improves embedding quality on MTEB, an embedding evaluation suite
All of this runs within a single unified model, avoiding the complexity and overhead of maintaining separate systems for retrieval and reranking.
Research Context#
The work addresses a practical need in search infrastructure, where computational efficiency and ranking quality both matter. By reconceptualizing reranking as a similarity operation over cached embeddings rather than a generative task, E2Rank opens a path for deployment in resource-constrained environments without sacrificing ranking performance.
