Reindex-Then-Adapt: Improving Large Language Models for Conversational Recommendation