/**
 * EntityExtractor — "discovery writer" (pantser) support
 *
 * Scans ALREADY-DRAFTED chapter prose and proposes NEW story entities
 * (characters and locations) that are not yet recorded in the project's
 * `characters/*.yml` / `locations/*.yml` files.
 *
 * This is a purely deterministic, heuristic analyser — there is NO LLM call.
 * The goal is to give a writer who "just writes" some structured value without
 * forcing them to define every entity up-front.
 *
 * Philosophy: under-propose rather than flood. Every threshold below is a
 * named constant and deliberately conservative — a missed candidate is cheap
 * (the writer can `create` it manually), whereas a wall of false positives
 * erodes trust in the command.
 */
/** A single proposed entity with how many times it appears in the chapter. */
export interface EntityCandidate {
    /** Canonical (capitalised) name as it appears in the prose. */
    name: string;
    /** Number of whole-word occurrences of `name` in the cleaned chapter text. */
    mentions: number;
}
/** Result of scanning one chapter for new entities. */
export interface ExtractionResult {
    /** Proposed NEW characters, sorted by mention count (descending). */
    characters: EntityCandidate[];
    /** Proposed NEW locations, sorted by mention count (descending). */
    locations: EntityCandidate[];
}
export declare class EntityExtractor {
    private readonly projectPath;
    constructor(projectPath: string);
    /**
     * Scan a single chapter file and propose NEW characters and locations.
     *
     * @param chapterFile - Absolute path, or a path relative to
     *                       `<projectPath>/chapters/`.
     */
    extractFromChapter(chapterFile: string): Promise<ExtractionResult>;
    /**
     * Scan an arbitrary prose file — typically a freeform outline or treatment the
     * author wrote before populating `characters/` / `plots/`. Use this to
     * bootstrap structured entities from an existing plan.
     *
     * @param file - Absolute path, or a path relative to `projectPath`.
     */
    extractFromFile(file: string): Promise<ExtractionResult>;
    /**
     * Core extraction over raw text (markup is stripped internally). Cross-checks
     * against the names already recorded under `characters/` and `locations/`.
     */
    extractFromText(raw: string): Promise<ExtractionResult>;
    /**
     * Build the character candidate list from two signals:
     *   1. Dialogue speakers identified via attribution patterns (strong signal —
     *      always proposed).
     *   2. Capitalised tokens that recur in non-sentence-start positions at least
     *      `MIN_CHARACTER_MENTIONS` times (proper-noun heuristic).
     *
     * Candidates already present in `known` (case-insensitive) are excluded.
     */
    private _extractCharacters;
    /**
     * Collect every dialogue speaker named via an attribution pattern.
     * Sentence-starter words (e.g. "Then said the man") are filtered out.
     */
    private static _collectSpeakers;
    /**
     * Collect capitalised single-word proper-noun candidates that recur in
     * NON-sentence-start positions at least `MIN_CHARACTER_MENTIONS` times.
     */
    private static _collectRecurringProperNouns;
    /**
     * Propose NEW locations: capitalised phrases that follow a locational
     * preposition. A candidate is accepted when it is either multi-word, ends in a
     * known place suffix, or is a single proper noun that is NOT already a
     * character candidate. Known locations are excluded.
     */
    private _extractLocations;
    /**
     * Load the canonical names already recorded under `<projectPath>/<subdir>/`.
     * Parses `name` and `fullName` from each `*.yml` / `*.yaml` file, and also
     * adds each individual word token so that e.g. an existing "Mara Quill"
     * suppresses a bare "Mara" candidate. All entries are lower-cased.
     *
     * File-based and best-effort: a missing directory or malformed YAML simply
     * yields fewer known names rather than throwing.
     */
    private _loadKnownNames;
    /**
     * Remove all non-prose markup so the heuristics only see narrative text:
     *   - YAML frontmatter (leading `--- … ---`)
     *   - HTML / scene comments (`<!-- … -->`)
     *   - `[MARKER: …]` annotations (TK, TODO, NOTE, …)
     *   - `--- Scene N ---` delimiters and bare horizontal rules
     *   - Markdown headings, emphasis markers, and link syntax
     */
    static stripMarkup(raw: string): string;
    /**
     * Split text into sentences for sentence-start detection. Splits on
     * end-of-sentence punctuation followed by whitespace, and on line breaks
     * (a new line conservatively begins a new "sentence" for capitalisation
     * purposes).
     */
    private static _splitSentences;
    /**
     * Count whole-word, case-insensitive occurrences of `term` in `text`.
     * Multi-word terms are matched as a contiguous phrase.
     */
    private static _countOccurrences;
}
//# sourceMappingURL=entity-extractor.d.ts.map