You built a RAG application in Azure AI Foundry that grounds a support chatbot on a large Azure AI Search index of product manuals. During evaluation, users report that the retrieved passages are often only loosely related to their questions, even though the correct answer exists somewhere in the index. Vector search recall is adequate, but the top-ranked chunks passed to the LLM are frequently not the most relevant ones. You want to improve the relevance of the passages that actually reach the model without re-chunking or re-embedding the entire corpus. What should you do?