--- name: web-research title: Research the web with web_search and web_fetch description: How to find and read web sources well with CLIO's web tools - writing queries, using filters, reading results and their saved full answers, recovering from failed reads, and quoting what a page actually says. keywords: - level:L0 - web --- `web_search` finds candidate pages; `web_fetch` reads one. A search result is a lead, not evidence: read the page before you rely on anything in it. ## Write the query - Name the thing precisely, with the terms an authoritative page would use (the standard's number, the library's name, the instrument, the agency). - Put constraints in the filters, not in the query text: `after`/`before` for dates (YYYY-MM-DD), `include_domains`/`exclude_domains` for sites, and `kind` (`research`, `news`, `pdf`) for the kind of page. - One idea per query. When results drift, rephrase with vocabulary taken from the best result so far, or search for the counter-claim. - For papers, prefer `literature_search` when the `literature` tools are available; use `web_search` for documentation, data portals, agency pages, software and grey literature. ## Read the results Each result has a title, URL, date when known and a short snippet. The `payload_file` holds everything the search source returned, including longer excerpts; read it with the file tools before running a near-identical search. `notes` explain how the answer was produced: - another source answered because the first was resting or used up; - `filters_applied_by_clio`: the source could not apply a filter, so CLIO removed results outside it. Results without a date stay in, so check dates yourself when they matter; - a source failed or needs the person's attention. These are facts about the search, not about the topic. Mention them only when they limit what you could find. A failed search says why in `error.advice`. A source that did not answer twice in a row (`search_source_temporarily_down`) has a temporary problem; it does not mean nothing matches. Call `web_search` again with the same query: the call waits `retry_after_s` and asks that source again, and if it fails again CLIO moves the search to `next_source`. Read pages you already found while you wait, if useful. ## Read a page `web_fetch` returns the page as Markdown in `content`, with the final `url` after redirects and a `title`. Cite the final URL. When a read fails: - try the same document at another address: the publisher's landing page, the arXiv `/abs/` or `/html/` page instead of a PDF, a mirror, or the DOI resolver; - a PDF is saved locally (`pdf_path`); read it with `view_pdf` when you can; - `reader` shows when a hosted service read the page for CLIO, which may serve a cached copy; prefer a fresh read when currency matters. ## Quote what the page says Copy the exact words from `content` when a claim depends on them, with the units, populations, dates and qualifications attached. Keep what the source states separate from what you infer from it.