/**
 * CompressionDetector
 *
 * Uses Normalized Compression Distance (NCD) to detect prompt injection
 * by measuring structural similarity between input and known attack templates.
 *
 * Technique: "Embedding similarity without embeddings"
 * - Uses Node.js built-in zlib (zero external dependencies)
 * - Compresses input concatenated with each attack template
 * - Similar strings compress more efficiently together → lower NCD score
 * - NCD ∈ [0, 1]: 0 = identical structure, 1 = maximally dissimilar
 *
 * Formula: NCD(x, y) = (C(xy) - min(C(x), C(y))) / max(C(x), C(y))
 *
 * Research basis:
 * - "Low-Resource Text Classification with Compressors" (ACL 2023)
 * - PromptGuard layered detection (Nature Scientific Reports 2025)
 *
 * Expected improvement: +3-5% F1 over regex-only detection
 */
export interface CompressionDetectorConfig {
    /** NCD threshold below which input is flagged as attack-similar (0-1, default: 0.55) */
    threshold?: number;
    /** Maximum number of templates to check per category (default: all) */
    maxTemplateChecks?: number;
    /** Custom attack templates to add to the built-in corpus */
    customTemplates?: Array<{
        category: string;
        template: string;
    }>;
    /** Minimum input length to analyze (default: 20) */
    minInputLength?: number;
    /** Maximum input length to analyze — truncates longer inputs (default: 2000) */
    maxInputLength?: number;
}
export interface CompressionDetectorResult {
    allowed: boolean;
    reason?: string;
    violations: string[];
    ncdAnalysis: {
        /** Lowest NCD score found across all templates */
        minNCD: number;
        /** Category of the closest-matching template */
        closestCategory: string;
        /** Average NCD across all checked templates */
        avgNCD: number;
        /** Number of templates checked */
        templatesChecked: number;
        /** Time taken in milliseconds */
        timeMs: number;
    };
}
/**
 * CompressionDetector — NCD-based prompt injection detection
 *
 * "Embedding similarity without embeddings" — uses gzip compression
 * to measure structural similarity between input and known attack templates.
 */
export declare class CompressionDetector {
    private config;
    private templates;
    constructor(config?: CompressionDetectorConfig);
    /**
     * Detect if input is structurally similar to known attack templates
     */
    detect(input: string): CompressionDetectorResult;
    /**
     * Compute NCD between input and a template
     * NCD(x, y) = (C(xy) - min(C(x), C(y))) / max(C(x), C(y))
     */
    private ncd;
    /**
     * Get compressed length of text using deflateRaw (no gzip header overhead)
     */
    private compressedLength;
    /**
     * Get the number of built-in templates
     */
    get templateCount(): number;
    /**
     * Get categories and their template counts
     */
    get categories(): Record<string, number>;
}