hwaaaaanii commited on
Commit
ff77331
·
verified ·
1 Parent(s): a35a7a9

Release RetailBridge benchmark

Browse files
.gitattributes CHANGED
@@ -58,3 +58,9 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
58
  # Video files - compressed
59
  *.mp4 filter=lfs diff=lfs merge=lfs -text
60
  *.webm filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
 
 
58
  # Video files - compressed
59
  *.mp4 filter=lfs diff=lfs merge=lfs -text
60
  *.webm filter=lfs diff=lfs merge=lfs -text
61
+ database/raw/inquiries.jsonl filter=lfs diff=lfs merge=lfs -text
62
+ database/raw/inquiry.sqlite filter=lfs diff=lfs merge=lfs -text
63
+ database/raw/order.sqlite filter=lfs diff=lfs merge=lfs -text
64
+ database/raw/product.sqlite filter=lfs diff=lfs merge=lfs -text
65
+ database/raw/review.sqlite filter=lfs diff=lfs merge=lfs -text
66
+ database/raw/reviews.jsonl filter=lfs diff=lfs merge=lfs -text
README.md ADDED
@@ -0,0 +1,218 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ language:
4
+ - en
5
+ pretty_name: RetailBridge
6
+ size_categories:
7
+ - 1K<n<10K
8
+ task_categories:
9
+ - question-answering
10
+ - table-question-answering
11
+ tags:
12
+ - retrieval-augmented-generation
13
+ - agentic-rag
14
+ - multi-hop
15
+ - text-to-sql
16
+ - heterogeneous-data
17
+ - retail
18
+ configs:
19
+ - config_name: queries
20
+ default: true
21
+ data_files:
22
+ - split: test
23
+ path: queries/test.parquet
24
+ - config_name: product
25
+ data_files:
26
+ - split: train
27
+ path: database/tables/product.parquet
28
+ - config_name: order
29
+ data_files:
30
+ - split: train
31
+ path: database/tables/order.parquet
32
+ - config_name: inquiry
33
+ data_files:
34
+ - split: train
35
+ path: database/tables/inquiry.parquet
36
+ - config_name: review
37
+ data_files:
38
+ - split: train
39
+ path: database/tables/review.parquet
40
+ - config_name: script
41
+ data_files:
42
+ - split: train
43
+ path: database/tables/script.parquet
44
+ - config_name: reviews_text
45
+ data_files:
46
+ - split: train
47
+ path: database/text/reviews.parquet
48
+ - config_name: inquiries_text
49
+ data_files:
50
+ - split: train
51
+ path: database/text/inquiries.parquet
52
+ - config_name: sales_scripts
53
+ data_files:
54
+ - split: train
55
+ path: database/text/sales_scripts.parquet
56
+ ---
57
+
58
+ # RetailBridge
59
+
60
+ **RetailBridge** is a multi-source question-answering benchmark for enterprise retail.
61
+ Each question is answered over a heterogeneous retail database: five SQL tables and
62
+ three free-text stores (reviews, customer inquiries, live-commerce sales scripts).
63
+ Evidence is spread across these sources, so answering a question means *bridging*
64
+ them: identifying the right sources, carrying entities and intermediate answers from
65
+ one source into the next, and producing an answer grounded in the retrieved records.
66
+
67
+ - **Paper:** [Toward Agentic Retrieval-Augmented Generation for Enterprise Retail Workflows over Heterogeneous Data Sources](https://arxiv.org/abs/2610.25096)
68
+ - **Code (RetailRAG evaluation framework):** [github.com/DISL-Lab/RetailBridge](https://github.com/DISL-Lab/RetailBridge)
69
+ - **Project page and leaderboard:** [retailbridge-kdd.github.io](https://retailbridge-kdd.github.io)
70
+
71
+ ## Contents
72
+
73
+ | Config | Rows | What it is |
74
+ |---|---|---|
75
+ | `queries` (default) | 1,680 | Benchmark questions with gold answers and evaluation labels |
76
+ | `product`, `order`, `inquiry`, `review`, `script` | 200 / 6,378 / 18,749 / 24,104 / 287 | SQL tables of the retail database |
77
+ | `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 |
78
+
79
+ The `database/raw/` folder holds the same database as the files the RetailRAG framework
80
+ reads directly (`*.sqlite` and `*.jsonl`).
81
+
82
+ ```python
83
+ from datasets import load_dataset
84
+
85
+ queries = load_dataset("DISLab/RetailBridge", "queries", split="test")
86
+ orders = load_dataset("DISLab/RetailBridge", "order", split="train")
87
+ reviews = load_dataset("DISLab/RetailBridge", "reviews_text", split="train")
88
+ ```
89
+
90
+ To run RetailRAG or another agent against the database, download the raw files:
91
+
92
+ ```python
93
+ from huggingface_hub import snapshot_download
94
+
95
+ path = snapshot_download("DISLab/RetailBridge", repo_type="dataset",
96
+ allow_patterns=["database/raw/*", "queries/*"])
97
+ ```
98
+
99
+ ## Benchmark questions
100
+
101
+ Questions are stratified along three evaluation axes.
102
+
103
+ | Axis | Field | Values |
104
+ |---|---|---|
105
+ | Reasoning capability | `reasoning_level` | 1, 2, 3 (number of hops) |
106
+ | Multi-source chaining | `source_chaining` | `Homo` (all hops use one source), `Hetero` (hops cross sources) |
107
+ | Evidence modality | `evidence_modality` | `sql-only`, `text-only`, `composite` (SQL and text) |
108
+
109
+ | Level | Chaining | SQL-only | Text-only | Composite | Total |
110
+ |---|---|---:|---:|---:|---:|
111
+ | 1 | Homo | 120 | 120 | – | 240 |
112
+ | 2 | Homo | 120 | 120 | – | 240 |
113
+ | 2 | Hetero | 120 | 120 | 240 | 480 |
114
+ | 3 | Homo | 120 | 120 | – | 240 |
115
+ | 3 | Hetero | 120 | 120 | 240 | 480 |
116
+ | **All** | | **600** | **600** | **480** | **1,680** |
117
+
118
+ Questions were generated by three models (DeepSeek-V4-Pro, Gemini-3.1-Pro, Qwen3.6-Max),
119
+ 560 each. Level-2 and Level-3 questions are composed from source-grounded single-hop
120
+ questions through two operators: **entity-linked** (a later hop is anchored on an entity
121
+ identified by an earlier hop) and **answer-linked** (a later hop takes an earlier hop's
122
+ answer as a value).
123
+
124
+ ### Fields
125
+
126
+ | Field | Type | Description |
127
+ |---|---|---|
128
+ | `id` | string | Question id |
129
+ | `question` | string | The question posed to the system |
130
+ | `answer` | string | Gold answer |
131
+ | `reasoning_level` | int | 1, 2 or 3 |
132
+ | `source_chaining` | string | `Homo` or `Hetero` |
133
+ | `sources` | list[string] | Gold evidence sources, in hop order; the reference for source-routing F1 (see *Source names*) |
134
+ | `evidence_modality` | string | `sql-only`, `text-only` or `composite` |
135
+ | `composite_family` | string or null | `entity_linked` or `answer_linked`; null for Level 1 |
136
+ | `operation_type` | string | Operation of the final hop: `search`, `aggregate`, `filter`, `rank`, `calculate`, `compare` |
137
+ | `generator` | string | Model that generated the question |
138
+
139
+ ### Example
140
+
141
+ ```json
142
+ {
143
+ "id": "ds_0000",
144
+ "question": "Which buyer has sent the most pre-purchase inquiries to Ashley X.?",
145
+ "answer": "Beth A.",
146
+ "reasoning_level": 1,
147
+ "source_chaining": "Homo",
148
+ "sources": ["inquiries"],
149
+ "evidence_modality": "sql-only",
150
+ "composite_family": null,
151
+ "operation_type": "rank",
152
+ "generator": "deepseek/deepseek-v4-pro"
153
+ }
154
+ ```
155
+
156
+ ## The database
157
+
158
+ The database comes from persona-driven simulations of a live-commerce retail platform
159
+ (product catalog, broadcast sales scripts, orders, pre- and post-purchase inquiries, and
160
+ reviews). Free text lives only in the text stores: the SQL tables carry ids, entities and
161
+ structured attributes, so any question that needs what a buyer or seller *said* has to go
162
+ through retrieval.
163
+
164
+ ### Source names
165
+
166
+ | Name in `sources` | Kind | Config | Raw file |
167
+ |---|---|---|---|
168
+ | `products` | SQL table `Product` | `product` | `product.sqlite` |
169
+ | `orders` | SQL table `Order` | `order` | `order.sqlite` |
170
+ | `inquiries` | SQL table `Inquiry` | `inquiry` | `inquiry.sqlite` |
171
+ | `reviews` | SQL table `Review` | `review` | `review.sqlite` |
172
+ | `scripts` | SQL table `Script` | `script` | `script.sqlite` |
173
+ | `inquiries_text` | Text store | `inquiries_text` | `inquiries.jsonl` |
174
+ | `reviews_text` | Text store | `reviews_text` | `reviews.jsonl` |
175
+ | `sales_scripts` | Text store | `sales_scripts` | `sales_scripts.jsonl` |
176
+
177
+ ### SQL tables
178
+
179
+ | Table | Columns |
180
+ |---|---|
181
+ | `Product` | product_id, product_name, product_category, original_price, store, features |
182
+ | `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 |
183
+ | `Inquiry` | inquiry_id, inquiry_type, buyer_id, buyer_name, seller_id, seller_name, product_id, product_name, script_id |
184
+ | `Review` | review_id, review_category, review_rating, buyer_id, buyer_name, seller_id, seller_name, product_id, product_name, script_id, inquiry_id |
185
+ | `Script` | script_id, seller_id, seller_name, product_id, product_name, broadcast_datetime, discount_rate, original_price, discount_price, discount_end_date, shipping_fee |
186
+
187
+ ### Text stores
188
+
189
+ Each document has `doc_id`, `content` (the text) and metadata columns that link it to
190
+ the SQL tables.
191
+
192
+ | Store | Content | Metadata |
193
+ |---|---|---|
194
+ | `reviews_text` | Review text | review_id, review_category, review_rating, buyer, seller and product ids/names, script_id, inquiry_id, order_id |
195
+ | `inquiries_text` | Buyer–seller inquiry dialogue | inquiry_id, inquiry_type, buyer, seller and product ids/names |
196
+ | `sales_scripts` | Live-commerce broadcast script | script_id, seller and product ids/names, broadcast_datetime, broadcast_id |
197
+
198
+ ## Evaluation
199
+
200
+ The [code repository](https://github.com/DISL-Lab/RetailBridge) contains RetailRAG and
201
+ the evaluation script. Answer accuracy is judged by an LLM (Muse Ember) against the gold
202
+ answer; source-routing F1 compares the sources an agent queries with `sources`; results for 16 LLM backbones are on the
203
+ [project page](https://retailbridge-kdd.github.io).
204
+
205
+ ## License
206
+
207
+ Apache-2.0.
208
+
209
+ ## Citation
210
+
211
+ ```bibtex
212
+ @article{choi2026retailbridge,
213
+ title = {Toward Agentic Retrieval-Augmented Generation for Enterprise Retail Workflows over Heterogeneous Data Sources},
214
+ author = {Choi, Jeonghwan and Hwang, Jibin and Park, Hayoon and Cheon, Hyeonjae and Song, Hwanjun},
215
+ journal = {arXiv preprint arXiv:2610.25096},
216
+ year = {2026}
217
+ }
218
+ ```
database/raw/inquiries.jsonl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fbb45e251e791c6208aa130d78c5c0337f61b312a9d249da8aa7cecdec519591
3
+ size 117209556
database/raw/inquiry.sqlite ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4c732757f5aa39225fd1862a7d631e8c463eeb5eda9a450ee277193ebd0ac8bb
3
+ size 3579904
database/raw/order.sqlite ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c49cb7974d43e1e2564b182c3c801da3fb4c94590efa5581d4b560234d19487f
3
+ size 1552384
database/raw/product.sqlite ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9258c613a1fa242c4a7262ada79490e6a9553d835d968fe82db745a8f04e6e3e
3
+ size 241664
database/raw/review.sqlite ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0d4c04ea5bdf625fbdfaa4042b99f44eb5ebda3199288c97e0c7d877f613a729
3
+ size 4984832
database/raw/reviews.jsonl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0497bf345acb74057cd29146acd9632f67955b29a530f5bdc379b0bdbcc2b448
3
+ size 17860292
database/raw/sales_scripts.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
database/raw/script.sqlite ADDED
Binary file (69.6 kB). View file
 
database/tables/inquiry.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:75dfab8bde4ccd330f726972bbf06d0794cdac5685d5579797f6256b2a355615
3
+ size 188313
database/tables/order.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:52acc57369d9b6a0495eb44786e27864841c155853371ef59f7b25b8ae800f38
3
+ size 116626
database/tables/product.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9fe367729a4cee1731ae572901def1b969b90c1d7d14efcfe2bdb3d3d456d6c0
3
+ size 131571
database/tables/review.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fbdf685485e9ce14e9c953c25e81333b9f95da0a1ad740c86a422e930dfaa5c1
3
+ size 321035
database/tables/script.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:48b9dd7fee9314775d3f704524802796893d6dcbaf5855be38c79bc9591a2814
3
+ size 35841
database/text/inquiries.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f7cfd4828a2e69f7f1ba378d27ace49c651477ac60fbf11155dd7c3c3a3e3440
3
+ size 53956455
database/text/reviews.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d2854351be8d1fa55424a05c27a2e69bd25f4c117e9ed3e414e14a4bc643586a
3
+ size 3927466
database/text/sales_scripts.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c284c5aa2559e5002a16e15a71365105eab86a68271701608a885234c42b4847
3
+ size 447175
queries/test.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
queries/test.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:581a22117adb5b59880a609c7f9510520f7b25d5acac116acfdb2900c98bae3a
3
+ size 312597