Use the web scraping tool to fetch full body text and align it into a table by specified fields—more honest than web search.
Disclaimer: The content shown on this page is generated by an AI model and is for reference only. It does not constitute any guarantee of effectiveness. Actual output and Credits consumption are subject to "Usage Details" in the console.
qwen3.7-plus (Responses API) → web_extractor + web_search (web scraping, fetch full body text) → align into a table by specified fields. This time the goal: extract the models on the Alibaba Cloud Model Studio "model catalog" page into a table.
One-sentence output: Turned the "model catalog" page into a 28-model × [model ID / modality / region / API protocol] comparison table, with zero fabricated model IDs, and it honestly states "region and protocol are inferred from URLs, the page doesn't state them directly"—expand for the full table it produced.
Steps: Use web scraping to build a table by field
The key is "use only the page's real text; if you can't find it, say so." Send the following instruction to qwen3.7-plus (web_extractor must be enabled):
View measured output · web_extractor scraping result as-is (excerpt, unmodified)
View measured output · web_extractor scraping result as-is (excerpt, unmodified)
Notes
- Consumption: This playbook involves text generation + built-in model tools.
web_extractormust be enabled together withweb_search; both are billed by successful call count, and the scraped page body enters the input as a whole. For this large page, the measuredweb_extractormade 4 calls and consumed about 88,000 total tokens; the larger the page and the more scraping rounds, the higher the input consumption. Validate on a small scale first. - Use web scraping, not web search: For the same type of task, switching to
web_searchwill have it assemble answers from fragmentary snippets and fill fields it can't find from memory (it has been measured fabricating brand names that don't appear on the page);web_extractorfetches the full page body, which is what yields this honest "extract what the page says" result—here all 28 model IDs genuinely exist on the page with zero fabrication (verified one by one against the original page). - Distinguish "extraction" from "inference": This time the model did not hard-write region/protocol as "not provided on the page"; instead it inferred them from real URLs on the page and explicitly stated the inference basis—this is acceptable, but you should be able to tell at a glance which parts are page text and which are inferred. The prompt line "if you can't find it, say so; do not fill in from memory" must be kept.