/**
 * Character-trigram text utilities for observational memory.
 *
 * Shared by the storage layer for:
 *   - building FTS5 MATCH queries against the trigram-tokenized index
 *     (`buildFtsMatchQuery`),
 *   - typo-tolerant search fallback (`bestTokenTrigramJaccard`),
 *   - write-time near-duplicate detection (`trigramJaccard`).
 *
 * Pure functions with no SQLite dependency so the logic is unit-testable
 * in isolation.
 */
/**
 * Minimum token length the SQLite trigram tokenizer can match. Tokens
 * shorter than this cannot be satisfied by an FTS5 MATCH query and must be
 * handled with complementary LIKE filtering (or the LIKE-only search path).
 */
export declare const MIN_TRIGRAM_TOKEN_LENGTH = 3;
/**
 * Per-token floor for {@link bestTokenTrigramJaccard}: every trigram-matchable
 * query token must reach this Jaccard similarity against some content token,
 * otherwise the content is considered not to match the query at all. This
 * mirrors the AND semantics of the FTS MATCH path — every term must be
 * (approximately) present — while still tolerating typos.
 */
export declare const TYPO_FALLBACK_MIN_TOKEN_SIMILARITY = 0.2;
/** Normalize text for trigram comparison: lowercase and collapse whitespace. */
export declare function normalizeForTrigrams(text: string): string;
/**
 * Build the set of character trigrams of `text`.
 *
 * Strings shorter than one trigram collapse to a single pseudo-token so
 * identical short strings still compare equal (Jaccard 1.0). Empty input
 * yields an empty set.
 */
export declare function trigramSet(text: string): Set<string>;
/**
 * Jaccard similarity between the character-trigram sets of two strings,
 * in [0, 1]. Returns 0 when either side has no trigrams.
 */
export declare function trigramJaccard(a: string, b: string): number;
/**
 * Typo-tolerant similarity between a whole query and a piece of content:
 * the mean, over query tokens, of the best Jaccard similarity between that
 * token and any single content token.
 *
 * Comparing token-to-token (instead of query-to-full-content) keeps the
 * score robust to sentence length — a one-word typo in a long observation
 * would otherwise be diluted below any useful threshold.
 *
 * Only `queryTokens` of trigram-matchable length are scored; shorter tokens
 * are ignored so they do not drag the mean down for matches they can never
 * express. Returns 0 — no match — when any such token fails to reach
 * TYPO_FALLBACK_MIN_TOKEN_SIMILARITY against every content token (the token
 * is absent, mirroring the AND semantics of the FTS path), or when no query
 * token qualifies.
 */
export declare function bestTokenTrigramJaccard(queryTokens: string[], content: string): number;
/**
 * Build an FTS5 MATCH query from search tokens.
 *
 * Each token becomes a quoted phrase (`"token"`); phrases are joined with
 * AND. With the trigram tokenizer each quoted phrase acts as a
 * case-insensitive substring match, so a multi-token query matches content
 * containing all tokens anywhere. Embedded double quotes are escaped by
 * doubling, which is how FTS5 string literals represent a literal quote.
 *
 * Returns an empty string when no non-empty tokens are supplied — callers
 * must guard against that before running a MATCH query.
 */
export declare function buildFtsMatchQuery(tokens: string[]): string;
//# sourceMappingURL=trigrams.d.ts.map