CAMI: Cost-Aware Agent-Guided Multi-Indexing for Semantic Retrieval

Abstract

Modern retrieval pipelines can benefit from augmentation of each corpus chunk with multiple semantic representation enrichments, such as content-conditioned hypothetical queries, summaries or paraphrases, to better align document representations with user queries. While these richer chunk representations can substantially improve retrieval quality, they also introduce a new systems challenge: the space of representation types and generator models is combinatorial, and the cost of exhaustive index-time evaluation scales linearly with corpus size. We present CAMI (Cost-Aware Multi-Indexing), a framework that treats multi-index construction as a multi-objective (cost vs accuracy) enrichment-portfolio selection problem. CAMI introduces three primary features: (i) an agentic discovery loop that leverages LLMs to propose usecase and corpus adaptive representation templates; (ii) a multi-objective search algorithm that combines independent statistical sampling with LLM-guided proposals to identify synergistic enrichment portfolio; and (iii) adaptive fidelity control with just-in-time enrichment, which prunes unpromising configurations via low-cost approximations while decoupling optimization spend from total corpus size. We evaluate CAMI across diverse RAG corpora, demonstrating that it automatically discovers high-recall representation portfolios over standard content-only baselines while operating under strict cost budgets. CAMI traces the recall–cost Pareto frontier providing a rigorous methodology for the end-to-end optimization of compound retrieval systems.