Retriever Portfolios: A Principled Approach to Adaptive RAG

Retrieval-augmented generation (RAG) systems typically rely on a single retriever and a single set of hyperparameters, despite facing highly heterogeneous queries that range from simple factoid questions to complex multi-hop reasoning. We propose a method that automatically selects a small, diverse subset of retrievers (a portfolio) from a large pool of candidates, to cover different regions of the target query distribution. We formalize this setting via an expected best-of-$k$ objective over the query distribution and show that it admits an efficient portfolio construction algorithm with near-optimal guarantees. Across multiple QA benchmarks, our learned portfolios and router pipeline consistently outperform single-retriever and naive multi-retriever baselines on both retrieval metrics and answer quality. In addition, compared to inference-time hyperparameter tuning approaches, fixed portfolios enable parallel retrieval and LLM calls, achieving comparable (and sometimes better) accuracy with substantially lower latency and token cost.

AI assistants that answer questions using documents usually work in two steps: first they search for relevant passages, then they generate an answer from those passages. Today, many systems use the same search method for every question, even though some questions need one obvious fact while others require combining evidence from several places. This paper asks whether an assistant can do better by keeping a small team, or portfolio, of search methods instead of betting on one. We introduce a principled way to choose this team from a large set of candidates, so that different members cover different kinds of questions. When a question arrives, a small decision model sends it to the most promising few search methods, whose answers can be produced in parallel. We also prove that this offline selection procedure comes close to the best possible team for the goal we set. Across several question-answering tests, these portfolios find better evidence and improve final answer accuracy compared with systems that use just one search method. They also avoid slow, repeated per-question tuning, making document-grounded AI systems more accurate and easier to deploy.