Datasets:
Improve dataset card readability
Browse files
README.md
CHANGED
|
@@ -58,11 +58,17 @@ configs:
|
|
| 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
|
| 62 |
three free-text stores (reviews, customer inquiries, live-commerce sales scripts).
|
| 63 |
-
|
| 64 |
-
|
| 65 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 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)
|
|
@@ -117,9 +123,10 @@ Questions are stratified along three evaluation axes.
|
|
| 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
|
| 121 |
-
|
| 122 |
-
|
|
|
|
| 123 |
|
| 124 |
### Fields
|
| 125 |
|
|
@@ -130,10 +137,10 @@ answer as a value).
|
|
| 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
|
| 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
|
| 137 |
| `generator` | string | Model that generated the question |
|
| 138 |
|
| 139 |
### Example
|
|
@@ -157,7 +164,7 @@ answer as a value).
|
|
| 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
|
| 161 |
structured attributes, so any question that needs what a buyer or seller *said* has to go
|
| 162 |
through retrieval.
|
| 163 |
|
|
|
|
| 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 with five SQL tables and
|
| 62 |
three free-text stores (reviews, customer inquiries, live-commerce sales scripts).
|
| 63 |
+
|
| 64 |
+
Evidence is spread across these sources, so answering a question means *bridging* them.
|
| 65 |
+
A system has to
|
| 66 |
+
|
| 67 |
+
- identify the right sources,
|
| 68 |
+
- carry entities and intermediate answers from one source into the next, and
|
| 69 |
+
- produce an answer grounded in the retrieved records.
|
| 70 |
+
|
| 71 |
+
## Links
|
| 72 |
|
| 73 |
- **Paper:** [Toward Agentic Retrieval-Augmented Generation for Enterprise Retail Workflows over Heterogeneous Data Sources](https://arxiv.org/abs/2610.25096)
|
| 74 |
- **Code (RetailRAG evaluation framework):** [github.com/DISL-Lab/RetailBridge](https://github.com/DISL-Lab/RetailBridge)
|
|
|
|
| 123 |
|
| 124 |
Questions were generated by three models (DeepSeek-V4-Pro, Gemini-3.1-Pro, Qwen3.6-Max),
|
| 125 |
560 each. Level-2 and Level-3 questions are composed from source-grounded single-hop
|
| 126 |
+
questions through two operators.
|
| 127 |
+
|
| 128 |
+
- **Entity-linked.** A later hop is anchored on an entity identified by an earlier hop.
|
| 129 |
+
- **Answer-linked.** A later hop takes an earlier hop's answer as a value.
|
| 130 |
|
| 131 |
### Fields
|
| 132 |
|
|
|
|
| 137 |
| `answer` | string | Gold answer |
|
| 138 |
| `reasoning_level` | int | 1, 2 or 3 |
|
| 139 |
| `source_chaining` | string | `Homo` or `Hetero` |
|
| 140 |
+
| `sources` | list[string] | Gold evidence sources in hop order, used as the reference for source-routing F1 (see *Source names*) |
|
| 141 |
| `evidence_modality` | string | `sql-only`, `text-only` or `composite` |
|
| 142 |
| `composite_family` | string or null | `entity_linked` or `answer_linked`; null for Level 1 |
|
| 143 |
+
| `operation_type` | string | Operation of the final hop (`search`, `aggregate`, `filter`, `rank`, `calculate`, `compare`) |
|
| 144 |
| `generator` | string | Model that generated the question |
|
| 145 |
|
| 146 |
### Example
|
|
|
|
| 164 |
|
| 165 |
The database comes from persona-driven simulations of a live-commerce retail platform
|
| 166 |
(product catalog, broadcast sales scripts, orders, pre- and post-purchase inquiries, and
|
| 167 |
+
reviews). Free text lives only in the text stores. The SQL tables carry ids, entities and
|
| 168 |
structured attributes, so any question that needs what a buyer or seller *said* has to go
|
| 169 |
through retrieval.
|
| 170 |
|