Wikexa Knowledge

The full text of an article, for when lookup()'s summary is not enough — sections as a JSON array, infobox as key/value facts, no HTML or wikitext to parse. Pass sections to pull only the parts you need (e.g. ["Early life"]) and max_chars to cap the payload; both exist because a long article will otherwise flood your context.

What a word means, in thousands of languages — 8.15M dictionary entries with senses, part of speech, etymology and pronunciation. Covers what a general model is weakest at: historical languages (Old English, Gothic, Ancient Greek, Middle French) and hundreds of regional and indigenous ones. A single spelling often has entries in many languages and you get all of them — hund returns Danish, Gothic, Icelandic, Middle English and more — or pass language to narrow, pos for one part of speech. Use this for words and lookup() for things: define("java") gives the word in eight languages, lookup("Java") gives the island.

Facts about any named thing — person, company, place, species, event, concept. Returns structured fields (dates, identifiers, relationships) plus a ~200-token summary, drawn from 10.2M entity records. Prefer this over fetching an encyclopedia page: the HTML costs ~15,000 tokens to recover ~500 tokens of fact. Resolves aliases and Wikidata Q-ids, so "Apple", "Apple Inc" and "Q312" all reach the same entity. Free, no key.

Academic paper metadata from 57M+ works — title, abstract, authors, citations, DOI and open access URL. Covers every field: CS, medicine, physics, economics, biology, and more. Search by keywords (e.g. "transformer attention"), by topic name (e.g. "machine learning"), or by OpenAlex topic ID. Returns papers sorted by citation count. Source: OpenAlex (CC0 metadata). Use this when the user needs scholarly references, citation counts, or research context that Wikipedia does not cover.

What changed in the last hours or days — the escape hatch for facts newer than your training cutoff. Reach for this whenever the answer could have moved since you were trained: elections, appointments, acquisitions, releases, deaths, records. Returns titles with timestamps and edit comments; resolve any of them with lookup(). Pass topic to filter and hours to widen the window up to a week.

Find the right title when you only have a partial name or a rough description. Returns ranked {title, wikidata_id, description, summary_snippet}; ranking blends text relevance with monthly pageviews and follows redirects, so abbreviations land on the real article — "usa" returns United States, "jfk" returns John F. Kennedy, "apple" returns Apple Inc. rather than a disambiguation page. Searches every corpus at once unless you pass corpus. Follow up with lookup() for facts or article() for the text.