What is the primary purpose of implementing a re-ranking technique in a retrieval pipeline?	To improve the accuracy of search results by reordering initially retrieved documents based on relevance to the user query.
Where does the re-ranking step fit into the overall retrieval process?	It is a post-processing step that occurs after the initial search results from the vector index and BM25 index have been merged.
What is the core task assigned to the Large Language Model (LLM) during the re-ranking process?	To review the user's question and the set of retrieved documents, and return the most relevant documents in order of decreasing relevance.
Why is it more efficient to ask the LLM to return Document IDs instead of the full text chunks?	Returning full text is computationally inefficient as the LLM must copy and process large amounts of data; returning IDs is a simple, lightweight instruction.
What technical methods are used to ensure the LLM returns well-formatted JSON during re-ranking?	Using an Assistant message pre-fill and a stop sequence.
What is the main advantage (pro) of adding a re-ranker to the search pipeline?	It significantly increases the accuracy of the search results.
What is the main disadvantage (con) of adding a re-ranker to the search pipeline?	It increases the latency of the search pipeline because it requires waiting for a call to the LLM.
What is the input provided to the re-ranker function?	The user's original query and the initial set of search results obtained from the vector and BM25 indices.