EPIC: Efficient Personalized Index Construction for Retrieval-Augmented Generation

Changmin Lee, Jaemin Kim, Taesik Gong

Keywords: Large Language Models, Retrieval-Augmented Generation, Personalization

TL;DR: Efficient Personalized Index Construction for Retrieval-Augmented Generation

Abstract: Personalizing retrieval-augmented generation (RAG) is a promising path for personal AI assistants but faces two key challenges: (i) indiscriminate indexing of large corpora imposes prohibitive memory costs, and (ii) preference-agnostic retrieval leads to mismatches with user preferences. We propose EPIC (Efficient Personalized Index Construction), a two-component framework that integrates preferences into both indexing and retrieval. EPIC performs preference-aware memory refinement, combining coarse-to-fine filtering and rewriting to build compact, preference-relevant memories, and preference-guided embedding steering, which adjusts query embeddings toward preference-aligned directions to improve retrieval fidelity. To benchmark personalization, we introduce three datasets—PrefWiki, PrefRQ, and PrefELI5—covering diverse domains and preferences. Experiments show that EPIC achieves the highest accuracy, the smallest memory footprint, and the lowest retrieval latency across baselines. Compared with the best-performing baseline, HippoRAG 2, EPIC improves accuracy by 10.1\%p, reduces indexing memory by 1110$\times$, and decreases latency by 110$\times$. As a plug-and-play module requiring no fine-tuning, EPIC enables efficient construction of preference-aligned memories for practical personalized RAG.

Primary Area: foundation or frontier models, including LLMs

Submission Number: 5761