import type { AtomicTextRange, SemanticChunker, TextChunk } from '../chunker/index.js';
import { type EmbedderInterface } from '../chunker/semantic-chunker.js';
import type { ParseResult } from '../parser/index.js';
import type { ImageRendition } from '../pdf-visual/types.js';
import { type SourceContext, type SourceMap } from '../utils/source-context.js';
import type { VectorChunk, VisualAttachment } from '../vectordb/index.js';
/** `embeddings` has the same length as `chunks`, index for index. */
export interface BuildChunksAndEmbeddingsResult {
    chunks: (TextChunk & {
        sourceContext?: SourceContext;
    })[];
    embeddings: number[][];
}
export interface BuildChunksFromParseResultResult extends BuildChunksAndEmbeddingsResult {
    visualAttachments: Map<number, VisualAttachment[]>;
    omittedImageCount: number;
}
export declare function findNearestChunk(chunks: readonly TextChunk[], sourceOffset: number): TextChunk | undefined;
export declare function createVisualAttachment(imageIndex: number, rendition: ImageRendition): VisualAttachment;
/**
 * Compute semantic chunks and their embeddings for already-extracted text.
 *
 * Does not fail fast on zero chunks — the MCP handler throws, the CLI warns
 * and returns 0. Chunker and embedder errors propagate verbatim.
 */
export declare function buildChunksAndEmbeddings(text: string, chunker: SemanticChunker, embedder: EmbedderInterface, options?: {
    sourceMap?: SourceMap | undefined;
    atomicRanges?: readonly AtomicTextRange[] | undefined;
    title?: string | null | undefined;
}): Promise<BuildChunksAndEmbeddingsResult>;
/**
 * Preserve the parser content/range mapping at one shared boundary. The title
 * affects embeddings only when the embedder opts into the title prefix.
 */
export declare function buildChunksFromParseResult(result: ParseResult, chunker: SemanticChunker, embedder: EmbedderInterface): Promise<BuildChunksFromParseResultResult>;
/**
 * Content identity of a source file: SHA-256 of its raw BYTES, not of the
 * parsed text, so any caller that can read the file reproduces it — which is
 * what lets a later sync decide "unchanged" without re-parsing.
 */
export declare function computeContentHash(bytes: Uint8Array): string;
/**
 * Build persistable `VectorChunk`s from computed chunks + embeddings — the
 * single chunk→row mapping for both the MCP handler and both CLI paths. One
 * shared `timestamp` per call, a fresh `id`, and `fileName`/`fileType` derived
 * with `node:path` so they hold on Windows.
 *
 * Throws when a chunk has no embedding; the two arrays must align 1:1.
 *
 * `contentHash` is `null` for a chunk set with no source file, and the key is
 * then omitted rather than stored empty, so a hashless row is never mistaken
 * for a real hash. Required, not optional, so a new call site cannot silently
 * write hashless rows.
 *
 * `visualProfile` is the requested visual intent of this ingestion, copied to
 * every row. It accepts `null`/`undefined` and then omits the optional property,
 * so a caller holding a nullable value can pass it directly under
 * `exactOptionalPropertyTypes` and a caller with no visual concept can omit it.
 */
export declare function buildVectorChunks(params: {
    filePath: string;
    chunks: (TextChunk & {
        sourceContext?: SourceContext;
    })[];
    embeddings: number[][];
    fileTitle: string | null;
    contentHash: string | null;
    visualAttachments?: ReadonlyMap<number, readonly VisualAttachment[]>;
    visualProfile?: string | null;
}): VectorChunk[];
//# sourceMappingURL=compute.d.ts.map