ecomagentbench/EcomAgentBenchProductDB

EComAgentBench Product Database

The prebuilt product and review database for EComAgentBench, accepted to EMNLP 2026 Industry Track.

Paper · Code and benchmark

This repository contains product.db, a SQLite database with FTS5 search indexes, covering approximately 3.7 million products and 21.4 million reviews. The file is 26.9GB.

The 662 benchmark tasks, agent code, and evaluation scripts are available in the GitHub repository.

Data source and use

The database is derived from Amazon Reviews 2023 by McAuley Lab, using the All_Beauty, Electronics, Cell_Phones_and_Accessories, and Office_Products categories.

It is released with permission from the dataset authors for research evaluation. Use is subject to the original dataset's terms.

Citation

If you use this database, please cite EComAgentBench and the underlying Amazon Reviews 2023 dataset.

@misc{du2026ecomagentbench,
  title         = {EComAgentBench: Benchmarking Shopping Agents on Long-Horizon Tasks with Distributed Hidden Intent},
  author        = {Zeyao Du and Tong Li and Haibo Zhang},
  year          = {2026},
  eprint        = {2606.17698},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2606.17698}
}