/**
 * Which language a prompt is written in, to the extent it can be told cheaply.
 *
 * **Why the estimator needs this at all.** Measured against the counting
 * endpoint, `estimateTokens` is accurate on English (+1.0%) and badly low on every
 * other Latin language: German -37.3%, Spanish -22.9%, French -15.1%. Characters
 * per token says why — English 3.44, French 2.66, Spanish 2.53, German 2.02 —
 * while the estimator applied one divisor to all of them. Non-English text is
 * thinner in the merge table, so the same number of characters costs more tokens.
 *
 * **Accents are not the signal.** That was the first hypothesis and it was tested
 * and killed: a Spanish sample with zero accented characters comes out at -22.9%,
 * against -22.1% for accented Spanish. Diacritics correlate with non-English text
 * in a corpus and not in a prompt, and weighting them moved the figure by three
 * points. What separates these languages is which words they are made of.
 *
 * So this counts function words — the shortest, commonest, most language-specific
 * tokens there are. `the of and to` against `der die und ist` against
 * `que de la en`.
 *
 * **It answers `null` when unsure, and that is the important part.** A prompt is
 * not an essay: it can be three lines, or English instructions wrapped around a
 * Spanish example, or a JSON schema with no prose at all. Guessing on those would
 * apply a language's divisor to text that is not in that language, which is how a
 * fix for one case becomes a regression for four. `null` means "use the default",
 * and the default is the English-calibrated behaviour this has always had.
 */
/** Languages this can name. Exported so callers can key a table off it. */
export declare const DETECTABLE_LANGUAGES: readonly string[];
/**
 * The language of `text`, or `null` when no answer is safe.
 *
 * Case-insensitive, whole words only, and it stops reading after a bounded
 * prefix: a prompt can be a megabyte, the answer does not get better after a few
 * thousand words, and this runs inside `estimateTokens` on every call.
 */
export declare function detectTextLanguage(text: string): string | null;
//# sourceMappingURL=language.d.ts.map