import type { Document as MupdfDocument } from 'mupdf';
import { type EmbedderInterface, type FilteredTextFragment } from './pdf-filter.js';
interface StextBbox {
    x: number;
    y: number;
    w: number;
    h: number;
}
/**
 * Shape of mupdf's structured-text JSON used by the per-page loop.
 * Captured here so both the items-extraction step and the raw `stextJson`
 * we return remain typed.
 */
interface StextJson {
    blocks: Array<{
        type: string;
        lines?: Array<{
            text: string;
            x: number;
            y: number;
            bbox?: StextBbox;
            font: {
                size: number;
                name?: string;
                weight?: string;
            };
        }>;
    }>;
}
/**
 * Per-page record produced by `extractPdfPages`. `text` is the page's text
 * after semantic header/footer filtering. `textFragments` carries every
 * survivor's native provenance and exact range in `text`; `stextJson` is the
 * copied raw mupdf structured-text JSON for downstream visual detection.
 */
interface ExtractedPage {
    pageNum: number;
    text: string;
    textFragments: FilteredTextFragment[];
    stextJson: StextJson;
}
/**
 * Result returned by `extractPdfPages`. The helper lifts three concerns
 * out of the legacy `parsePdf` body:
 *   1. the per-page `toStructuredText` + `block.type === 'text'` loop;
 *   2. `filterPageBoundarySentences` for header/footer removal;
 *   3. title-resolution materials (`metadataTitle` and `page1FontHint`).
 *
 * Both `parsePdf` and `parsePdfPages` consume this helper; they differ only
 * in the `stextOptions` argument they pass to `page.toStructuredText(...)`.
 */
interface ExtractedPdf {
    pages: ExtractedPage[];
    metadataTitle: string | undefined;
    page1FontHint: {
        text: string;
        fontSize: number;
    } | undefined;
}
/**
 * Per-page extraction shared by `parsePdf` and `parsePdfPages`.
 *
 * Takes an already-open mupdf `Document` and:
 *   - reads `info:Title` once,
 *   - iterates pages calling `toStructuredText(stextOptions)`,
 *   - builds `PageData` items (only `block.type === 'text'` lines),
 *   - runs `filterPageBoundarySentences` to drop semantic headers/footers,
 *   - derives `page1FontHint` from page 1's largest-font lines.
 *
 * The two callers differ ONLY in `stextOptions`: `parsePdf` passes
 * `'preserve-whitespace'`;
 * `parsePdfPages` passes `'preserve-whitespace,preserve-images'` so mupdf
 * emits `block.type === 'image'` entries for the downstream visual-candidate
 * detector.
 *
 * Lifecycle: this helper does NOT call `doc.destroy()` — disposal stays
 * with the caller.
 */
export declare function extractPdfPages(doc: MupdfDocument, embedder: EmbedderInterface, stextOptions: string): Promise<ExtractedPdf>;
export {};
//# sourceMappingURL=pdf-extract.d.ts.map