ItemRAG: Item-Level Retrieval for LLM-Based Recommendation
A new research paper proposes ItemRAG, a retrieval-augmented generation (RAG) approach for large language model (LLM)-based recommender systems. Unlike existing methods that retrieve purchase histories of similar users—often noisy or weakly relevant—ItemRAG shifts to fine-grained item-level retrieval. It augments descriptions of items in a target user's history or candidate set by retrieving items relevant to each. The goal is to provide more informative context for LLMs to recommend items, especially for cold-start scenarios. The paper is available on arXiv under ID 2511.15141.
Key facts
- ItemRAG is a novel RAG approach for LLM-based recommendation.
- It shifts from user-history retrieval to item-level retrieval.
- It augments item descriptions by retrieving relevant items.
- Aims to reduce noise and improve recommendation accuracy.
- Addresses cold-start items using LLM reasoning.
- Published on arXiv with ID 2511.15141.
- The paper is a preprint (replace-cross).
- Focuses on improving context for candidate items.
Entities
Institutions
- arXiv