/**
 * File chunking for semantic indexing.
 *
 * Splits a file into overlapping line windows, each capped by characters so a
 * chunk stays within the embedding model's effective token window (MiniLM
 * truncates around 256 tokens ≈ 1000 chars). Chunk ids are `relpath#index`;
 * the id → line-range mapping is kept in the sidecar metadata (index-meta.ts)
 * so search hits can be rendered as `path:start-end`.
 *
 * Bump CHUNKER_VERSION when the strategy changes — a version mismatch in the
 * sidecar triggers a clean rebuild of the store.
 */
export declare const CHUNKER_VERSION = 2;
/**
 * Hard character cap per chunk.
 *
 * The "~256 tokens ≈ 1000 chars" this was set from does not hold for code:
 * measured against the bundled tokenizer over this repo, 1000 chars is **313
 * tokens** at the median, so MiniLM (256) truncates 85.8% of chunks and drops
 * 20.3% of the corpus's tokens, while bge-small (512) drops 0.1%. Raising this
 * is therefore not free in the way the old comment implied — it spends a
 * budget that is already overdrawn on one model and nearly full on the other.
 */
export declare const CHUNK_MAX_CHARS = 1000;
export interface Chunk {
    /** `relpath#index` — the id stored in the vector index. */
    id: string;
    /** Text sent to the embedder. */
    text: string;
    /** 1-based inclusive start line. */
    startLine: number;
    /** 1-based inclusive end line. */
    endLine: number;
}
/**
 * Split `content` into chunks. `relPath` becomes the id prefix. Returns an
 * empty array for empty or binary-looking content.
 *
 * `maxChars` overrides {@link CHUNK_MAX_CHARS} for eval arms that sweep the
 * window. Production never passes it; a caller that does is changing what the
 * index contains and owes the store a distinct key, since nothing about a
 * stored vector records the cap it was built under.
 */
export declare function chunkFile(relPath: string, content: string, maxChars?: number): Chunk[];
//# sourceMappingURL=chunker.d.ts.map