information-retrieval

Fetch a page from a tollbooth-registered domain. Costs the publisher's fee_credits from your credit balance (top up at /credits/buy) plus this route's price; 80% of the fee accrues to the publisher. Returns extracted page content. 404 if the domain isn't registered; 409 (not 402) if your credit balance is short.

AI model capability, specs and limits table for 2,000+ LLMs — GPT-5, GPT-4o, Claude 5, Claude Sonnet, Gemini 3 Pro, Llama 4, DeepSeek V4 and more: context window size, max output tokens, vision/audio/function-calling/reasoning support. Compare model constraints side-by-side across every provider to pick the right model for long-context, multimodal, or tool-use workloads. Refreshed hourly.

Borrower health-factor lookup by EVM address. Returns current health factor, collateral, debt, and protocol across all monitored lending markets.

Ask a database in plain English (text to SQL): any BigQuery table, including Google's public datasets. The SQL is written for you, run under a byte ceiling, and the rows are read back into a written answer, returned with the SQL, columns and rows it used. Use it for analytics questions without writing SQL. Input: question, table; optional columns.

Deep web research with citations: search the live web for your question, read the top sources in full, and get a written answer with every claim tied to a numbered source. Use it when an agent needs a sourced answer, not a list of links. Input: question; optional max_sources. Quick answer with links: search.web.

Fact check claims against sources: up to 10 claims checked against up to 4 source URLs you name; each comes back SUPPORTED, CONTRADICTED or UNSUPPORTED with the exact quote the verdict rests on. Use it to verify AI output, citations or statements before publishing. Input: claims[], sources[]. A question but no sources yet: research.web.

Screen a name against the five official sanctions lists (EU, US OFAC SDN, UN, UK FCDO, French asset freezes) from a snapshot refreshed daily, with per-list dates. Query: name=NAME (optional threshold=0.6-1.0, lists=eu,ofac,un,uk,fr). Returns scored matches with entity type and programs, and a verdict: match, possible_match or no_match.

Fetch a URL and extract exactly the JSON you ask for. Pass a field `schema` (or `instructions`) and get a validated JSON object pulled from the page content.

Holt die genannten Quellen, zerlegt sie in Saetze und sucht die Fundstellen mit der hoechsten Begriffsdeckung zur Behauptung. Liefert woertliche Zitate mit Quellenangabe, erkennt Verneinung und prueft, ob die Zahlen aus der Behauptung in den Quellen vorkommen. Bewertet, was in den Quellen steht, nicht was zutrifft.

Structured data extraction / text to JSON: pull the fields you name out of any text into a JSON object with exactly those keys, null where the text does not state a value (Gemini, no invented values). Use it to turn emails, invoices, pages or documents into clean records. Input: text, fields[]. From a live web page: research.page_facts.

Run Python code in a sandbox: execute Python in Google's hosted sandbox (not on this service's machines) and get the code, its printed output and the outcome back. Use it for calculations, data transforms and checks an agent should not do in its head. Input: code. Questions over BigQuery data: data.query.

Compare a page against its last stored state, or two supplied documents, and report only the content changes.

Matches declared classification tokens against a supplied published dual-use control list. Missing facts are returned as null and named in unavailable_fields where applicable. Derived values disclose their method; source-backed values disclose source, time window and age. This is neither legal advice nor customs advice.

Retrieves current package metadata from npm, PyPI, crates.io or Maven Central. Missing facts are returned as null and named in unavailable_fields where applicable. Derived values disclose their method; source-backed values disclose source, time window and age.

Website change monitoring, step one: fetch a page and store it as the baseline to compare against later. Use it to watch pricing pages, terms, job boards or competitors. Input: url. Then call monitor.check to learn whether and how it changed.

BigQuery SQL: run a read-only SQL query, including against Google's public datasets (Wikipedia, GitHub, blockchains, weather, census and more), and get columns and rows back as JSON with the bytes processed. Every query is dry-run first and refused above the scan ceiling, so cost is bounded before it runs. Input: sql; optional max_scan_gib. Ask in plain English instead: data.question.

Website change detection, step two: re-fetch a page saved with monitor.snapshot and get whether it changed, how old the baseline is, and a written summary of what changed. Use it to track price changes, terms updates, new job posts or competitor moves. Input: url; optional language.

Loest die vollstaendige Weiterleitungskette auf und belegt jeden Sprung mit Statuscode, Zielhost, Protokoll-, Registrierungsdomain-, Land- und ASN-Wechsel. Loest Kurz-URLs bis Tiefe fuenf auf, meldet Punycode und Homographen in beiden Schreibweisen, prueft Zertifikat, Domainalter, Registrar, Nameserver, ASN, Hoster und Land, fragt offene Sperrlisten und DNSBL fuer die Ziel-IP ab, vergleicht Content-Type gegen Dateiendung und durchsucht den Seitentext nach versteckten Anweisungen an Sprachmodelle

A monitor with memory over registered trials. The first call stores a snapshot and returns the baseline explicitly marked as such rather than reporting everything as new. Each later call with the returned token compares the current registry state and reports the differences: recruitment status, enrolment target, primary and overall completion dates, the number of sites, and whether results have been posted.

Fetches the feed over https with redirect and size limits, recognises RSS 2.0, RDF and Atom, and converts each entry into a NIP-01 event. Kind 1 produces a short note built from title, optional shortened summary and the link; kind 30023 produces a long-form article carrying the full entry text plus the title, published_at and d tags that NIP-23 expects.

Validates output against a JSON Schema and reports each violation with its JSON path, the rule it broke and a sentence saying what to change. Accepts either a parsed object or the raw string a model returned: a payload wrapped in a code fence or surrounded by prose is unwrapped, and the wrapping is reported separately from schema errors so you fix the right layer. Draft 2020-12 and draft-07 are both supported, along with the common string formats.

Prueft Fremdtext vor der Verarbeitung durch ein Sprachmodell auf zwoelf Angriffsmuster in Deutsch und Englisch, auf unsichtbare Steuerzeichen, auf den Unicode-Tag-Bereich als Schmuggelkanal und auf Base64-Bloecke, deren dekodierter Inhalt selbst wieder Anweisungen enthaelt. Liefert Risikowert, Fundstellen mit Auszug und einen bereinigten Text.

Findet E-Mail, Telefon, IBAN mit Pruefsumme, Kreditkarte mit Luhn-Pruefung, IP-Adressen, Steuer- und Umsatzsteuernummern, Kennzeichen, Sozialversicherungsnummern und Geburtsdaten mit Kontextwort. Liefert je Treffer Kategorie, Zeile, Spalte und Position sowie den geschwaerzten Text in drei Schwaerzungsarten. Die Fundwerte werden nicht gespeichert.

Calculates item co-occurrence and item similarity from the submitted events only. No user profile is stored.