/**
 * Markdown-aware section chunker for the chunk-level search index
 * (`fs_blob_chunks`). Pure text → chunks; no SQL, no model calls.
 *
 * Design (card #2, from the chunking research on card #1):
 *  - The chunk unit is a heading-bounded markdown section; a document with no
 *    headings that fits the budget is a single whole-page chunk.
 *  - Oversized sections split at paragraph boundaries, and a paragraph that
 *    alone exceeds the budget splits at line boundaries — never mid-line, so
 *    every chunk is an exact line slice of the source (`startLine..endLine`,
 *    1-indexed inclusive) and can be re-read from the file.
 *  - Zero overlap between chunks; adjacency is recoverable from line ranges.
 *  - Each body chunk's indexed `content` is prefixed with its breadcrumb
 *    (document title > heading path) so a chunk that is just a price table
 *    still embeds/searches with its context. The prefix counts against the
 *    token budget (metadata-aware budgeting).
 *  - YAML front matter (`---` fenced block at line 1) becomes chunk 0 as-is
 *    (minus any `volatileFrontmatterKeys` in the indexed content): a
 *    whole-document signal chunk carrying title/summary/keyword fields.
 */
export interface ChunkingOptions {
    /**
     * Token budget per chunk, breadcrumb prefix included. Default 480 — under
     * the 512-token input ceiling of the strictest multilingual embedding
     * models, with headroom for tokenizer variance across languages.
     */
    maxTokens?: number;
    /**
     * Token estimator. Default: `ceil(length / 4)` — the usual chars-per-token
     * heuristic; inject a real tokenizer for exact budgets.
     */
    estimateTokens?: (text: string) => number;
    /**
     * Top-level front-matter keys to strip from chunk 0's indexed `content`
     * (the `body` and line range stay exact). For keys that change on every
     * write without carrying meaning — crawl timestamps like `fetchedAt` —
     * this keeps chunk 0's content hash stable, so embedding caches keyed by
     * content survive re-crawls. A stripped key's indented continuation lines
     * (multi-line values, lists) are stripped with it.
     */
    volatileFrontmatterKeys?: string[];
}
export interface MarkdownChunk {
    /** Position of this chunk in the document, 0-based and gapless. */
    index: number;
    /** First source line of the chunk, 1-indexed, inclusive. */
    startLine: number;
    /** Last source line of the chunk, 1-indexed, inclusive. */
    endLine: number;
    /** Breadcrumb of the enclosing headings ("Title > H2 > H3"), null when
     *  the document gives no title/heading context (or for front matter). */
    headingPath: string | null;
    /** Exact slice of the source: lines `startLine..endLine` joined by `\n`. */
    body: string;
    /** What gets indexed/embedded: the breadcrumb prefix + the body (for the
     *  front-matter chunk: the body minus any `volatileFrontmatterKeys`). */
    content: string;
}
/**
 * Chunk one markdown document. Returns chunks in document order with exact,
 * non-overlapping 1-indexed line ranges. Empty/blank input yields no chunks.
 */
export declare function chunkMarkdown(text: string, options?: ChunkingOptions): MarkdownChunk[];
//# sourceMappingURL=chunking.d.ts.map