An LLM as arbiter in RAG retrieval: picking the right candidate with reasons

Instead of just combining retrieval scores, a more advanced RAG system uses a large language model as a final arbiter to select the best answer candidates. This arbiter is given a structured brief for each candidate, detailing which retrieval methods found it, the document section it's in, and a snippet of the text. In a single, efficient call, the LLM then assigns each candidate a role, such as "primary" or "supporting," and provides a clear, human-readable reason for its decision. This method is superior to traditional score fusion because it understands the rich context behind why a passage was retrieved, leading to more defensible and accurate results.