RCEM preprint reports conversational search gains while preserving existing indexes

Researchers propose a conversational search method called RCEM that reports up to 30% improvement over prior approaches under distributional shift, when evaluation data differs from training data. The preprint describes a way to handle conversational queries while retaining the original search model's embedding space, allowing existing passage indexes to remain in use without rebuilding them. RCEM trains the model to align conversations with queries rewritten by a large language model rather than directly with relevant passages. This removes the need for conversation-to-passage relevance labels during training. The reported improvement is an upper bound from experiments under distributional shift.