Iskat

Search the web index

Search the Webdex index of crawled websites. Combines keyword (BM25) and semantic (vector) retrieval. Supports "quoted phrases" for exact matches and -term to exclude a term. Works across languages: a query in one language can retrieve pages written in another.

Find products with prices

Search shop catalogues by name and specification, filter by price and availability, and sort by price. Use this rather than `search` whenever the question is what something costs or where to buy it: `search` returns pages that mention a product, this returns the offer itself — price, stock, specifications and a link straight to the product page. Specifications are searchable, so a query like "ноутбук 16 гб 512 ссд" or "холодильник samsung no frost" matches on capacity and features, not just the

Read an indexed page

Return the full extracted text of a page already in the index. This serves stored content and does not hit the live site, so it is fast and cannot be blocked. Use it after search to read a result.

Read several indexed pages at once

Like fetch_page but for up to 20 URLs in one call — the bulk-read counterpart to search_batch, so reading a page of results is one round trip instead of ten. Each page comes back with its cleaned text from the index; URLs not in the index are marked, not fatal.

Find pages similar to a page

Given an indexed URL, return other indexed pages with similar content, compared by meaning rather than shared keywords.

Run several searches at once

Execute up to 10 queries in one call. Answering a real question usually takes several related searches, and doing them one at a time spends a round trip on each; this runs them together and returns the results grouped by query.

Get quotable passages, not just links

Return the specific passages that answer a question, each with the URL and title it came from. Use this when you intend to cite or summarise rather than browse: it skips the step of fetching pages and hunting for the relevant paragraph, and the attribution is exact.

What kind of source is this

Describe a site in the index: what type of source it is (official, media, reference, forum, shop), how much of it we hold, how trusted it is by the link graph, and how fresh the copy is. Use this before relying on a result. A ministry's own page and a forum post repeating it are both 'a URL', and the difference decides how much weight the claim deserves.

Track a query for new results

Save a query and later ask what has appeared since you last checked. Pull-based: nothing is pushed to you, you call check_watch when you want to know — which fits an agent that runs on its own schedule. Use for monitoring: a topic, a competitor, a regulation, a product.

What is new for a tracked query

Return pages matching a watched query that were indexed since the last check, then advance the cursor. Call with no id to check all of your watches at once.

List tracked queries

Show the queries you are watching and when each was last checked.

Stop tracking a query

Remove a watch. Pass its id (from list_watches) or the exact query text. Watches accumulated with no way to remove them; this closes that.

List indexed sites

Show which sites the index covers, with page counts. Use this to understand coverage before concluding that something is missing.

Index statistics

Size and freshness of the corpus: sites, pages, passages, languages.