Zerank-1 as a Cross-Encoder (Cross-Encoder) is significantly different from the traditional embedding model (Bi-Encoder) in terms of architecture and application:
-
Differences in treatment: the normal embedding model generates separate vector representations for the query and the document respectively, and then computes the similarity between these two vectors; whereas Zerank-1 processes the entire content of the query and the document at the same time for deeper interaction analysis.
-
Precision vs. efficiency trade-offs: Cross-encoders typically provide higher sorting accuracy because they capture the complex interactions between queries and documents; however, this architecture requires more computations to be performed and is therefore slower to process, making it suitable for use as a second-stage fine-grained sorter.
-
Different application scenarios: The common embedding model is suitable for handling initial retrieval of large number of documents; while Zerank-1 is suitable for fine ranking of a small number of candidates (e.g., 100-1000). Practical systems often use a combination of the two techniques: fast recall by the embedding model, followed by precise reordering by Zerank-1.
This technical difference makes Zerank-1 particularly suitable for scenarios requiring high accuracy, such as enterprise-level search, RAG systems and intelligent Q&A applications.
This answer comes from the articleZerank-1: A reordering model for improving the precision of search resultsThe
