RetailBridge / README.md
hwaaaaanii's picture
Note the KDD 2027 acceptance and clarify the sources field
d40f339 verified
|
Raw History Blame Contribute Delete
8.92 kB
metadata
license: apache-2.0
language:
  - en
pretty_name: RetailBridge
size_categories:
  - 1K<n<10K
task_categories:
  - question-answering
  - table-question-answering
tags:
  - retrieval-augmented-generation
  - agentic-rag
  - multi-hop
  - text-to-sql
  - heterogeneous-data
  - retail
configs:
  - config_name: queries
    default: true
    data_files:
      - split: test
        path: queries/test.parquet
  - config_name: product
    data_files:
      - split: train
        path: database/tables/product.parquet
  - config_name: order
    data_files:
      - split: train
        path: database/tables/order.parquet
  - config_name: inquiry
    data_files:
      - split: train
        path: database/tables/inquiry.parquet
  - config_name: review
    data_files:
      - split: train
        path: database/tables/review.parquet
  - config_name: script
    data_files:
      - split: train
        path: database/tables/script.parquet
  - config_name: reviews_text
    data_files:
      - split: train
        path: database/text/reviews.parquet
  - config_name: inquiries_text
    data_files:
      - split: train
        path: database/text/inquiries.parquet
  - config_name: sales_scripts
    data_files:
      - split: train
        path: database/text/sales_scripts.parquet

RetailBridge

Accepted at KDD 2027 (Datasets and Benchmarks Track).

RetailBridge is a multi-source question-answering benchmark for enterprise retail. Each question is answered over a heterogeneous retail database with five SQL tables and three free-text stores (reviews, customer inquiries, live-commerce sales scripts).

Evidence is spread across these sources, so answering a question means bridging them. A system has to

  • identify the right sources,
  • carry entities and intermediate answers from one source into the next, and
  • produce an answer grounded in the retrieved records.

Links

Contents

Config Rows What it is
queries (default) 1,680 Benchmark questions with gold answers and evaluation labels
product, order, inquiry, review, script 200 / 6,378 / 18,749 / 24,104 / 287 SQL tables of the retail database
reviews_text, inquiries_text, sales_scripts 24,104 / 18,749 / 287 Free-text stores (review text, buyer–seller inquiry dialogues, live-commerce sales scripts), each with metadata

Usage

Download the whole benchmark

One call downloads the questions and the full heterogeneous database. The database/raw/ folder holds the five SQL tables as *.sqlite files and the three text stores as *.jsonl files, which is the form the RetailRAG framework reads directly.

from huggingface_hub import snapshot_download

path = snapshot_download("DISLab/RetailBridge", repo_type="dataset")
# path/queries/test.jsonl         1,680 questions
# path/database/raw/*.sqlite      5 SQL tables
# path/database/raw/*.jsonl       3 text stores

Load with datasets

Each source is a separate config because the tables have different columns. The loop below loads the questions and every source at once.

from datasets import load_dataset

REPO = "DISLab/RetailBridge"
queries = load_dataset(REPO, "queries", split="test")
db = {name: load_dataset(REPO, name, split="train")
      for name in ["product", "order", "inquiry", "review", "script",
                   "reviews_text", "inquiries_text", "sales_scripts"]}

Benchmark questions

Questions are stratified along three evaluation axes.

Axis Field Values
Reasoning capability reasoning_level 1, 2, 3 (number of hops)
Multi-source chaining source_chaining Homo (all hops use one source), Hetero (hops cross sources)
Evidence modality evidence_modality sql-only, text-only, composite (SQL and text)
Level Chaining SQL-only Text-only Composite Total
1 Homo 120 120 – 240
2 Homo 120 120 – 240
2 Hetero 120 120 240 480
3 Homo 120 120 – 240
3 Hetero 120 120 240 480
All 600 600 480 1,680

Questions were generated by three models (DeepSeek-V4-Pro, Gemini-3.1-Pro, Qwen3.6-Max), 560 each. Level-2 and Level-3 questions are composed from source-grounded single-hop questions through two operators.

  • Entity-linked. A later hop is anchored on an entity identified by an earlier hop.
  • Answer-linked. A later hop takes an earlier hop's answer as a value.

Fields

Field Type Description
id string Question id
question string The question posed to the system
answer string Gold answer
reasoning_level int 1, 2 or 3
source_chaining string Homo or Hetero
sources list[string] Gold evidence sources in hop order (see Source names)
evidence_modality string sql-only, text-only or composite
composite_family string or null entity_linked or answer_linked; null for Level 1
operation_type string Operation of the final hop (search, aggregate, filter, rank, calculate, compare)
generator string Model that generated the question

Example

{
  "id": "ds_0000",
  "question": "Which buyer has sent the most pre-purchase inquiries to Ashley X.?",
  "answer": "Beth A.",
  "reasoning_level": 1,
  "source_chaining": "Homo",
  "sources": ["inquiries"],
  "evidence_modality": "sql-only",
  "composite_family": null,
  "operation_type": "rank",
  "generator": "deepseek/deepseek-v4-pro"
}

The database

The database comes from persona-driven simulations of a live-commerce retail platform (product catalog, broadcast sales scripts, orders, pre- and post-purchase inquiries, and reviews). Free text lives only in the text stores. The SQL tables carry ids, entities and structured attributes, so any question that needs what a buyer or seller said has to go through retrieval.

Source names

Name in sources Kind Config Raw file
products SQL table Product product product.sqlite
orders SQL table Order order order.sqlite
inquiries SQL table Inquiry inquiry inquiry.sqlite
reviews SQL table Review review review.sqlite
scripts SQL table Script script script.sqlite
inquiries_text Text store inquiries_text inquiries.jsonl
reviews_text Text store reviews_text reviews.jsonl
sales_scripts Text store sales_scripts sales_scripts.jsonl

SQL tables

Table Columns
Product product_id, product_name, product_category, original_price, store, features
Order order_id, script_id, buyer_id, buyer_name, product_id, product_name, quantity, total_price, status, order_date, delivery_date, post_inquiry_type, seller_id, seller_name
Inquiry inquiry_id, inquiry_type, buyer_id, buyer_name, seller_id, seller_name, product_id, product_name, script_id
Review review_id, review_category, review_rating, buyer_id, buyer_name, seller_id, seller_name, product_id, product_name, script_id, inquiry_id
Script script_id, seller_id, seller_name, product_id, product_name, broadcast_datetime, discount_rate, original_price, discount_price, discount_end_date, shipping_fee

Text stores

Each document has doc_id, content (the text) and metadata columns that link it to the SQL tables.

Store Content Metadata
reviews_text Review text review_id, review_category, review_rating, buyer, seller and product ids/names, script_id, inquiry_id, order_id
inquiries_text Buyer–seller inquiry dialogue inquiry_id, inquiry_type, buyer, seller and product ids/names
sales_scripts Live-commerce broadcast script script_id, seller and product ids/names, broadcast_datetime, broadcast_id

Evaluation

The code repository contains RetailRAG and the evaluation script. Answer accuracy is judged by an LLM (Muse Ember) against the gold answer; results for 16 LLM backbones are on the project page.

License

Apache-2.0.

Citation

@article{choi2026retailbridge,
  title   = {Toward Agentic Retrieval-Augmented Generation for Enterprise Retail Workflows over Heterogeneous Data Sources},
  author  = {Choi, Jeonghwan and Hwang, Jibin and Park, Hayoon and Cheon, Hyeonjae and Song, Hwanjun},
  journal = {arXiv preprint arXiv:2610.25096},
  year    = {2026}
}