Personalized Retrieval-Augmented Generation For Human-AI Collaborative Data Analysis

Abstract

Retrieval-Augmented Generation (RAG) has become the dominant approach for grounding large language model outputs in verified, domain-specific knowledge, reducing hallucinated responses by over 60 percent relative to non-retrieval baselines [12]. Current RAG systems for enterprise analytics, however, treat every user identically: a senior data engineer and a business analyst asking the same question receive the same retrieval, ranking, and generated response, despite needing fundamentally different things. This article argues that personalization is not an optional refinement but a structural requirement for analytics RAG, since the right answer in data analysis depends on who is asking, what they already know, and what they intend to do with the result. We propose a personalized RAG architecture built on three components: a User Profile Store combining explicit and implicit signals, a Multi-Index Retrieval System routing queries across specialized schema, metric-definition, past-query, documentation, and example indices, and a Personalized Generation Layer that adapts response format and depth to the individual analyst. Drawing on personalized-search literature showing 10 to 31 percent MAP improvements from personalized re-ranking [14], and on the author's own experience building a comparable multi-domain retrieval-augmented analytics assistant in production, we argue this architecture is both theoretically grounded and practically achievable. We close with a proposed evaluation methodology for personalized analytics RAG, an area with no existing standardized benchmark.