Cost-Aware Query Routing in RAG: Empirical Analysis of Retrieval Depth Tradeoffs

Collection of recent IR/ML papers (arXiv June 3, 2026) focused on retrieval-augmented generation, dense retrieval, streaming ANNS, fairness/bias, and practical systems. Key contributions include CA-RAG: a per-query routing framework that trades retrieval depth, latency, token cost, and quality (FAISS + OpenAI evaluation, reproducible CSVs); attention calibration to reduce positional bias in dense retrievers; evidence that supervised neural retrievers learn query-independent document relevance priors causing findability gaps; Slipstream: a locality-aware method to speed up streaming ANNS (Faiss