| --- |
| task_categories: |
| - text-generation |
| language: |
| - en |
| tags: |
| - web-search |
| - search-api |
| - code |
| - agents |
| - tool-use |
| - enterprise-api |
| - sap |
| - servicenow |
| - workday |
| size_categories: |
| - n<1K |
| configs: |
| - config_name: hard_tickets |
| data_files: |
| - split: eval |
| path: data/eval.jsonl |
| --- |
| |
| # Web Search for Coding Agents: Validation Set |
|
|
| #### 75 realistic product tickets for developing and tuning against the **web search providers** benchmark for coding agents. The model, task, budgets, and runner stay fixed; the search or fetch provider is the variable under test. |
|
|
|
|
| ## Dataset Description |
|
|
| 75 examples: Each row comes from a **distinct source URL**. Each row has: |
|
|
| * `task_id`: stable identifier, prefixed by vendor. |
| * `vendor`: The vendor the question is about. |
| * `ticket`: the product request in natural language. It deliberately does not |
| name the answer. |
| * `starter_code`: a real but wrong API call. It would run today; it does not |
| satisfy the ticket. |
| * `gold_code`: reference patch, used for grading. The model never sees it. |
|
|
| The rows follow the same format as the public set |
| ([openbenchmarks/OB-Code-Websearch](https://huggingface.co/datasets/openbenchmarks/OB-Code-Websearch)) |
| and the held-out private set, so a harness built against them runs unchanged on |
| the scored set. The vendor mix follows the private set, which leans toward |
| ServiceNow and Workday. Harness, scoring and vendor runners are in |
| [openbenchmarks-labs/web-search-for-coding-agents](https://github.com/openbenchmarks-labs/web-search-for-coding-agents). |
|
|
| ## How the tickets were built |
|
|
| 1. **Discover.** Official documentation pages for each vendor: the ServiceNow |
| REST API reference, Workday's published OpenAPI specs, and SAP pages found |
| through web search on SAP's official sites. Pages off those sites are dropped. |
| 2. **Fetch.** The full text of each page is saved locally. |
| 3. **Generate.** An authoring model reads only those saved pages and proposes |
| tickets: an opaque gold token copied from the page, capability English that |
| does not name it, and a starter that calls a real but wrong API. |
| 4. **Verify.** An automatic keep-gate rejects a ticket that names its gold token, |
| whose gold does not compile, or whose starter already contains the answer. |
| Every gold patch passes the grader and every starter fails it. |
| 5. **Memory check.** Each ticket was run with no search tools. Tickets the model |
| solved from memory were removed. |
|
|
| ## Dataset design note |
|
|
| Vendors with highly regular, developer-friendly documentation can be poor fits |
| for this benchmark. In earlier Stripe tasks, the model often inferred the |
| documentation slug or URL structure directly from the ticket and reached the |
| correct implementation without performing meaningful retrieval. When |
| documentation followed a predictable pattern such as |
| `/changelog/dahlia/{kebab-case-ticket-title}`, the model could construct the |
| likely documentation location from the ticket itself. Those tasks measured how |
| well a model exploits documentation naming conventions, which is a different |
| skill from search quality, so they were removed from the active benchmark. |
|
|
|
|
