RetailBridgeRead the paper ↗
KDD 2027 · DATASETS AND BENCHMARKS

RetailBridge

Toward Agentic Retrieval-Augmented Generation for Enterprise Retail Workflows over Heterogeneous Data Sources

Jeonghwan Choi1,*Jibin Hwang1,*Hayoon Park1Hyeonjae Cheon3Hwanjun Song1,2,†
1 Korea Advanced Institute of Science and Technology (KAIST)2 Cluvion3 GS Retail

* Equal contribution   ·   † Corresponding author

PaperCode Repo
The benchmark at a glance
1,680verified QA pairs
8

Evidence sources

3

Reasoning levels

16

LLM backbones

01 / MOTIVATION

Why must retail agents bridge sources?

Enterprise retail workflows require agents to chain evidence across heterogeneous sources, from structured product and order records to unstructured inquiries and reviews.

THE GAP

Yet real enterprise data is rarely accessible, and existing agentic RAG benchmarks rarely capture these cross-source dependencies.

THIS WORK

We introduce RetailBridge, a compositional benchmark built from persona-driven retail simulations, and RetailRAG, a retail-aware framework for evaluating LLM agents on it.

What an agent must do
  1. 1

    Identify the relevant sources

    Decide which tables and text collections hold the evidence a question needs.

  2. 2

    Carry evidence between them

    Pass entities and intermediate answers from one source into the next.

  3. 3

    Produce a grounded answer

    Return an answer backed by the records that were actually retrieved.

02 / RESULTSBenchmarking LLMs on RetailBridge

Which LLMs can bridge evidence between sources in a heterogeneous retail database?

We evaluate 16 LLM backbones on RetailBridge using the same RetailRAG pipeline.

Answer accuracy
LLM-as-judge (Muse Ember)
Source routing F1
queried vs. gold sources
Efficiency
mean API calls per query

Answer accuracy

HIGHER IS BETTER ↑

Loading published results…

Answer accuracy vs. cost per queryWhich backbones give the most accuracy per dollar?

UP AND LEFT IS BETTER ↖
Open-sourceProprietaryDashed ring: estimated cost

Answer accuracy: LLM-as-judge (Muse Ember), overall. Cost per query: mean USD per question on OpenRouter (log scale), covering every LLM call made while answering; it excludes the one-time schema briefing, retrieval embeddings, and judge calls.

03 / FRAMEWORK

RetailBridgeMulti-Source QA over Heterogeneous Retail Sources

Figure 1: Overview of RetailBridge. Starting from persona-driven retail simulations, RetailBridge constructs a heterogeneous retail database, then (Step 1) generates source-grounded atomic QAs and (Step 2) recursively composes them via entity-linked and answer-linked operators into multi-source, multi-hop queries stratified along three evaluation axes.

Alongside RetailBridge, we propose RetailRAG as a retail-aware reference framework for evaluating LLMs over heterogeneous retail sources.

RetailRAGA Retail-Aware Agentic RAG Framework

Figure 2: Overview of RetailRAG, a retail-aware agentic RAG built on (i) auto-schema briefing, (ii) source-routed plan chaining, and (iii) diagnostic plan revision.
04 / BENCHMARK
1,680Total QA pairs
Generators

Qwen3.6-MaxDeepSeek-V4-ProGemini-3.1-Pro

Validators

GPT-5.5Claude Opus 4.7Human annotators

01

Reasoning Capability

One, two, or three hops reveal where compositional reasoning breaks down.

  • Level-124014.3%
  • Level-272042.9%
  • Level-372042.9%
02

Multi-Source Chaining

Single-source and multi-source queries isolate the challenge of bridging evidence.

  • Homo72042.9%
  • Hetero96057.1%
03

Evidence Modality

SQL, text, and combined evidence test both retrieval and cross-modality reasoning.

  • SQL-only60035.7%
  • Text-only60035.7%
  • Composite48028.6%
WORKED EXAMPLE FROM RETAILBRIDGELevel 3 · Text + SQL · 3 sources

What is the average original price of the products purchased from Robert A. by the buyer who gave a 5-star review for Robert A.'s Libby's Mixed Vegetables, mentioning their 'soup vibe had just passed'?

  1. 1
    Find the reviewerHOP 1 · TEXT RETRIEVAL
    REVIEWS_TEXT · POST-CS · 5 / 5
    The support person was so sweet and totally understood that my soup vibe had just passed, which made the whole return feel really easy. I love that they were so kind and didn't make me feel weird about my change of heart!
    BUYER IDENTIFIEDStanley M.
  2. 2
    List the purchasesHOP 2 · SQL QUERY
    ORDERS · STRUCTURED RECORDS
    SELECT DISTINCT product_name
    FROM "Order"
    WHERE seller_name = 'Robert A.'
      AND buyer_name = 'Stanley M.'
    QUERY RESULT · 4 PRODUCTS
    • Libby's Mixed Vegetables …
    • Jelly Belly 20-Flavor Spring Gift Box, 8.5-oz
    • SUNMALL 821-2049-A HDD Cable …
    • FLYRUN Men's Athletic Ankle Quarter Socks …
  3. 3
    Average the pricesHOP 3 · SQL QUERY
    PRODUCTS · STRUCTURED RECORDS
    SELECT AVG(original_price)
    FROM Product
    WHERE product_name IN (
      'Libby''s Mixed …',
      'Jelly Belly …',
      'SUNMALL …',
      'FLYRUN …')
    BENCHMARK GOLD ANSWER15.9
05 / BIBTEX

Cite This Work

@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}
}