SearchX

Search

Find pages about a topic. Returns ranked results with titles, URLs, snippets and trust scores — snippets, not full text. Skills: search first, then web_fetch the one or two URLs worth reading; a snippet is rarely enough to answer from. Use site to stay inside one domain ('lex.uz'), freshness for recent-only ('week', or an ISO range), language when the answer is language-specific. Ask for 3-5 results, not 20: more results cost context and rarely add an answer. If you already know what sentence y

Answer

Ask a question and get sources for it, plus a synthesized answer where a language model is available. Skills: this deployment currently has no language model, so 'answer' comes back empty and you get the sources only — read them with web_fetch, or use web_lookup instead, which is built for finding the sentence that answers a question. Check the warnings field: it says plainly when the model was unavailable rather than pretending the page had nothing to say.

Web Fetch

Read one URL as clean markdown, with navigation, ads and boilerplate removed. JavaScript-heavy pages are rendered in a browser automatically. Skills: this is how you actually read a page — search gives you the address, web_fetch gives you the text. Default max_length is 10000 characters, which truncates long pages: set max_length=0 for a full legal code or article, or chunked=true to page through it with chunk_index instead of holding it all. When you need one provision or fact from a long page

Extract Document

Extract text from a document you hold: PDF, Word, Excel, PowerPoint, OpenDocument, RTF, EPUB or CSV, sent as base64. Scanned pages are read with OCR. Skills: send the bytes once. For anything book-sized, send it with chunk_size and you get back a map — how many chunks, what each is about, and content_hash — then read one page at a time with content_hash + chunk_index, never resending the file. Follow next_chunk_index until has_more is false; jump straight to a chunk whose title matches what you

Web Lookup

Search, open the top results in parallel, and return only the passages containing your term. One call instead of search-then-fetch-each. Skills: reach for this when you want a specific fact, quote or provision rather than an overview — it is the cheapest way to answer a question, because it returns passages instead of whole pages. Set find to the exact word you expect on the page when it differs from the query ('volumes' while querying 'docker compose'); it matches on word boundaries. Raise top

Suggest

Autocomplete a partial query. Skills: for building a query box, not for research — it returns query strings, never answers or URLs. Needs at least 2 characters. If you are trying to discover what a site calls something, search and read the titles instead. This endpoint requires an API key.