/**
 * Librarian chunker — LLM-powered fact extraction for free-form text.
 *
 * Accepts a block of raw text about a single entity and returns a list of
 * structured `EntryInput` facts ready to be written to the knowledge store.
 * Uses the router's `extraction` task to pick a fast/cheap model.
 *
 * Extraction strategy:
 *  1. Primary pass — send the full `rawContent` to the LLM and parse the JSON
 *     array of `{ key, value, summary, confidence }` facts.
 *  2. Sentence-level fallback — if the primary pass returns zero facts AND the
 *     text splits into multiple sentences/paragraphs, retry each segment
 *     individually to capture facts that the full-context prompt missed.
 *  3. Deduplication — merge segments by key, keeping the first occurrence.
 *
 * Confidence blending: extracted confidence (0-100) is blended 70/30 with the
 * caller's `input.confidence` so that low-confidence ingestion paths can't be
 * inflated by an overconfident LLM extraction.
 */
import { EntryInput, EntityType } from '../types';
export interface ChunkInput {
    entityType: EntityType;
    entityId: string;
    rawContent: string;
    source: string;
    confidence: number;
    createdBy: string;
}
export interface ChunkResult {
    chunks: EntryInput[];
    extractedCandidates: number;
    skipped: number;
    reason?: string;
}
export declare function chunkContent(input: ChunkInput): Promise<ChunkResult>;
//# sourceMappingURL=chunker.d.ts.map