KDD 2026 Tutorial

Retrieval-Augmented Generation for LLM-based Recommender Systems

A comprehensive tutorial on retrieval, augmentation, and generation strategies that ground large language models in external knowledge for better recommendations.

August 09–13, 2026 Jeju Island, Republic of Korea ~3 hours (with break)
View Schedule Meet the Speakers
01 — Overview

The First Tutorial on RAG for LLM-based RecSys

Large language models are increasingly adopted for recommendation tasks thanks to their reasoning capabilities and effectiveness with cold-start items. This tutorial presents a comprehensive taxonomy of retrieval-augmented generation techniques tailored for LLM-based recommender systems — covering the types of input information, design choices at each RAG pipeline stage (retrieval, augmentation, generation), and open research questions.

Target audience: researchers and practitioners interested in recommender systems, information retrieval, and large language models. Basic familiarity with RecSys and LLMs is helpful but not required.

02 — Motivation

Why RAG for Recommendation?

LLMs offer strong reasoning and generation abilities, but often lack personalized, up-to-date, and domain-specific knowledge. RAG bridges this gap by grounding recommendations in retrieved user, item, and contextual information.

🔍

How can retrieval improve LLM-based recommendation quality?

📦

What should be retrieved — users, items, reviews, knowledge graphs, or demonstrations?

🧩

How should retrieved information be represented and injected into LLMs?

📊

How can we evaluate RAG-based recommender systems effectively?

03 — Schedule

Tutorial Outline

A structured walk-through of the RAG pipeline for recommendation, from fundamentals to open problems.

20 min Part I — Introduction
  • Preliminaries on LLM-based recommender systems
  • Preliminaries on retrieval-augmented generation (RAG)
  • Advantages and key considerations for adapting RAG to LLM-based RecSys
30 min Part II — Input Information
  • Target user information: interaction-based, query-based, and biography-based
  • External databases: interaction-related and item-related
40 min Part III — Retrieval Strategies
  • Interaction-based retrieval (user-item interactions)
  • Semantic-based retrieval (textual / visual similarity)
  • Knowledge graph-based retrieval (entities, paths, subgraphs)
  • Other strategies: web-based & SQL-based retrieval
20 min ☕ Break
30 min Part IV — Augmentation Strategies
  • Task prompt augmentation (auxiliary users, enriched user info)
  • Candidate prompt augmentation (candidate items, enriched descriptions)
  • Token-level augmentation (soft prompts from vector representations)
20 min Part V — Generation Strategies
  • Single-LLM generation
  • Multi-LLM generation with different roles
20 min Part VI — Open Questions & Conclusion
  • User privacy in user-retrieval-based approaches
  • Token efficiency for faster recommendation serving
  • Diversity of retrieved information for diverse results
04 — Speakers

Tutorial Presenters

A team of researchers from KAIST and Snap Inc. with deep expertise in recommender systems, graph mining, and LLMs.

Sunwoo Kim

Sunwoo Kim*

Ph.D. Student, KAIST

Research on aligning foundation models for recommendation, graph representation learning. Best Survey Paper at PAKDD 2025. Published at KDD, WWW, NeurIPS, ICLR, ICML, EMNLP, CVPR.

Website Email
Geon Lee

Geon Lee*

Post-doctoral Researcher, KAIST

Research on graph-based and LLM-based recommendation, complex interaction systems. Best Survey Paper at PAKDD 2025, Best Paper at ICDM 2025. Published at WWW, RecSys, KDD, NeurIPS.

Website Email
Kyungho Kim

Kyungho Kim*

Ph.D. Student, KAIST

Research on GNN and LLM-based recommender systems. Best Survey Paper at PAKDD 2025. Published at WWW, CIKM, ACL, AAAI, NeurIPS, RecSys.

Website Email
Liam Collins

Liam Collins

Research Scientist, Snap Inc.

Research on recommendation systems, user modeling, and language modeling. Best Resource Paper at CIKM 2025. Published at NeurIPS, KDD, WSDM, SIGIR, ICML.

Website Email
Neil Shah

Neil Shah

Principal Scientist, Snap Inc.

Large-scale representation learning, recommender systems, and efficient ML. 80+ publications, 8800+ citations. Best paper awards at KDD and CHI.

Website Email
Kijung Shin

Kijung Shin

Associate Professor, KAIST (Corresponding)

Data mining and ML for graph-structured data. 100+ papers, 4 best-paper awards. Organized 8 tutorials at KDD, WWW, ICDM, CIKM, AAAI.

Website Email

* Equal contribution

05 — Materials

Tutorial Resources

📑

Slides

Full tutorial slide deck
📄

Survey Paper

Comprehensive survey on RAG for LLM-based RecSys
💻

Code Repository

Implementations and examples
📚

Reading List

Key references and further reading
06 — Open Questions

Directions for Future Research

01
User Privacy — How can user-retrieval-based approaches preserve privacy while still retrieving meaningful collaborative signals?
02
Token Efficiency — How can we reduce token usage in RAG pipelines for faster, cost-effective recommendation serving?
03
Diversity — How can we improve the diversity of retrieved information to produce more diverse recommendation results?