ItemRAG

This repository is an official implementation of ItemRAG, an item-based retrieval-augmented generation technique for LLM-based recommendation.

Paper information

• Title: ItemRAG: Item-based Retrieval-Augmented Generation for LLM-based Recommendation.
• Authos: Sunwoo Kim, Geon Lee, Kyungho Kim, Jaemin Yoo, and Kijung Shin.
• Venue: SIGIR 2026 (short paper)

Datasets

In this work, we use the four datasets from https://amazon-reviews-2023.github.io/.

Name	# Users	# Items
Sports & Outdoors	25,363	15,701
Toys & Games	19,026	14,718
Beauty & Personal Care	45,490	31,151
Arts, Crafts & Sewing	24,511	18,884

The entire dataset, including (1) user-item interactions and (3) item titles, is presented in the link below:

• Link: https://www.dropbox.com/scl/fo/olpz13hyfcdn5jg6a4tzy/AN1Na5w_ySyO4nnYFyYnRpE?rlkey=5i6iyloeq9fpa48tz29ztmaqj&st=ibnulph8&dl=0
• Description: Refer to the README.txt file within the link.

Running code

ItemRAG consists of the four steps below:

• Step 1: Find semantically similar items based on the textual similarity, which is implemented in step1_get_similar_title_items.py
• Step 2: Retrieve relevant items, which is implemented in step2_do_retrieval.py
• Step 3: Generate a summary of the relevant items, which is implemented in step3_generate_summary.py
• Step 4: Perform final LLM-based recommendation, which is implemented in evaluation.py

The overall repository hierarchy is expected to be formed as follows:

/
 - dataset
   - /dataset/sports_outdoors_full.pickle
   - /dataset/toys_games_full.pickle
   - /dataset/beauty_care_full.pickle
   - /dataset/arts_full.pickle
 - step1_get_similar_title_items.py
 - step2_do_retrieval.py
 - step3_generate_summary.py
 - evaluation.py

All files include a term --dataset. This indicates the name of the target dataset. It should be one of:

• Sports & Outdoors -> 'sports_outdoors'
• Toys & Games -> 'toys_games'
• Beauty & Personal Care -> 'beauty_care'
• Arts, Crafts & Sewing -> 'arts'

We detail each Python file:

• Step 1: Finding semantically similar items. '--device' indicates the backbone GPU device for a language model. '--K' indicates the number of semantically similar items for each.

python3 step1_get_similar_title_items.py --dataset sports_outdoors --device cuda:0 --K 5

• Step 2: Retrieving relevant items. '--K' indicates the number of items to retrieve.

python3 step2_do_retrieval.py --dataset sports_outdoors --K 50

• Step 3: Generating summary. Note that one should put a proper api key for GPT in the file!

python3 step3_generate_summary.py --dataset sports_outdoors

• Step 4: Performing LLM-based recommendation.

python3 evaluation.py --dataset sports_outdoors