import { TrainingExportAdapter, TrainingExportRecord } from '@remnic/core';

/**
 * WeClone Alpaca-format training export adapter.
 *
 * Converts TrainingExportRecord[] into the JSON format that
 * WeClone / LLaMA Factory expects for fine-tuning:
 *
 *   [{ "instruction": "...", "input": "", "output": "..." }, ...]
 *
 * Only the three Alpaca fields are emitted; Remnic-specific
 * metadata (category, confidence, sourceIds) is stripped.
 */

declare const wecloneExportAdapter: TrainingExportAdapter;

/**
 * Communication style marker extraction.
 *
 * Analyzes text samples using simple heuristics to produce
 * a StyleMarkers profile.  No LLM calls — pure regex and
 * counting.
 */
interface StyleMarkers {
    avgSentenceLength: number;
    usesEmoji: boolean;
    formality: "formal" | "casual" | "mixed";
    usesLowercase: boolean;
    commonPhrases: string[];
}
/**
 * Analyse text samples and extract communication style markers.
 */
declare function extractStyleMarkers(samples: string[]): StyleMarkers;

/**
 * Training-pair synthesizer.
 *
 * Converts Remnic's flat TrainingExportRecord[] — where
 * `instruction` is a natural-language description and
 * `category` identifies the memory type — into natural
 * conversational question-answer pairs suitable for
 * WeClone / LLaMA Factory fine-tuning.
 *
 * Uses template-based question generation (no LLM calls).
 */

interface SynthesizerOptions {
    styleMarkers?: StyleMarkers;
    maxPairsPerRecord?: number;
}
/**
 * Synthesize natural conversational training pairs from
 * category-tagged memory records.
 */
declare function synthesizeTrainingPairs(records: TrainingExportRecord[], options?: SynthesizerOptions): TrainingExportRecord[];

/**
 * PII privacy sweep for training export records.
 *
 * Belt-and-suspenders check that runs after Remnic's own
 * privacy controls.  Scans instruction, input, and output
 * fields for common PII patterns and replaces matches with
 * [REDACTED].
 */

interface PrivacySweepResult {
    cleanRecords: TrainingExportRecord[];
    redactedCount: number;
    redactionDetails: {
        index: number;
        field: string;
        pattern: string;
    }[];
}
/**
 * Scan and redact PII from training export records.
 *
 * Returns a new array of cleaned records, leaving the originals
 * unmodified.  The `redactedCount` is the number of records that
 * had at least one redaction.  `redactionDetails` lists every
 * individual match with its record index, field, and pattern name.
 */
declare function sweepPii(records: TrainingExportRecord[]): PrivacySweepResult;

/**
 * @remnic/export-weclone
 *
 * WeClone-specific training-data export adapter that converts
 * Remnic memories into Alpaca-format fine-tuning datasets
 * compatible with WeClone / LLaMA Factory.
 */

interface TrainingExportRegistry {
    getTrainingExportAdapter(name: string): TrainingExportAdapter | undefined;
    registerTrainingExportAdapter(adapter: TrainingExportAdapter): void;
}
/**
 * Idempotently register the WeClone adapter with the core training-export
 * registry. Callable multiple times without throwing (CLAUDE.md #13:
 * secondary calls must not crash host processes that pre-register the
 * adapter for test fixtures).
 *
 * Returns true when the adapter was newly registered, false when an adapter
 * with the same name already exists.
 */
declare function ensureWecloneExportAdapterRegistered(registry?: TrainingExportRegistry): boolean;

export { type PrivacySweepResult, type StyleMarkers, type SynthesizerOptions, type TrainingExportRegistry, ensureWecloneExportAdapterRegistered, extractStyleMarkers, sweepPii, synthesizeTrainingPairs, wecloneExportAdapter };