A research team has published CAMI, a framework designed to make retrieval-augmented generation pipelines cheaper to configure without sacrificing recall quality.
RAG systems often layer extra indices on top of a base document index — think synthetic queries or auto-generated summaries — to close the gap between how documents are written and how users phrase their questions. The problem: deciding which enrichment types to generate, and with which models, creates a combinatorial explosion. The cost of testing every option scales with corpus size, so production teams either overspend or guess. CAMI reframes this as a budget-constrained portfolio selection problem. It uses an agentic discovery phase to propose enrichment templates specific to the corpus, then searches for combinations that work well together, pruning losing configurations early so the search cost stays flat regardless of how large the corpus grows.
The results are worth noting: CAMI hit up to 9.4% better recall@10 than content-only baselines in hard retrieval settings, and found those winning portfolios using up to five times less compute than random search. For teams running large-scale RAG in production, that efficiency gap is the difference between a tuning step you can afford and one you skip.
RAG optimization has mostly been a retrieval-time problem — rerankers, query rewriting, hybrid search. CAMI pushes the lever to index construction time, which is less glamorous but arguably more durable; a better index helps every query, not just the ones a reranker catches.