This research introduces compound retrieval systems that optimize how multiple ranking models interact beyond traditional cascading approaches.

• Proposes a more flexible framework where models collaborate rather than simply filtering in stages
• Balances ranking quality with computational costs through optimized model interaction
• Enables more efficient document retrieval while maintaining result quality
• Provides a mathematical framework for systematically optimizing multi-model systems

For linguists, this work offers new ways to leverage language models in search applications, enabling more nuanced relevance assessment while managing computational constraints.

Optimizing Compound Retrieval Systems