/**
 * LLM Prompting Framework for POWL Model Generation
 *
 * Implementation of prompting strategies from:
 * "Process Modeling With Large Language Models" (Kourani et al., 2024)
 *
 * Four key strategies:
 * 1. Role Prompting - Assign LLM the role of process modeling expert
 * 2. Knowledge Injection - Provide POWL language knowledge
 * 3. Few-Shot Learning - Provide input/output examples
 * 4. Negative Prompting - Specify what to avoid
 */
/**
 * Message format for LLM conversation history
 */
export interface LLMMessage {
    role: "system" | "user" | "assistant";
    content: string;
}
/**
 * Prompt templates for LLM-based POWL generation
 */
export declare const PROMPT_TEMPLATES: {
    /**
     * System prompt combining role assignment + knowledge injection
     *
     * Strategy 1 (Role) + Strategy 2 (Knowledge Injection)
     */
    SYSTEM_PROMPT: string;
    /**
     * Few-shot examples from the Kourani et al. paper
     *
     * Strategy 3: Few-Shot Learning
     */
    EXAMPLES: string;
    /**
     * Negative prompting - common errors to avoid
     *
     * Strategy 4: Negative Prompting
     */
    COMMON_ERRORS: string;
    /**
     * Generate a complete prompt for process model generation
     *
     * @param processDescription Natural language description of the process
     * @param conversationHistory Previous messages in the conversation (for error refinement)
     * @returns Complete prompt as a user message
     */
    generatePrompt(processDescription: string, conversationHistory?: LLMMessage[]): string;
    /**
     * Generate error refinement prompt
     *
     * @param error The error message or validation failure
     * @param conversationHistory Full conversation history
     * @returns Refinement prompt message
     */
    ERROR_REFINEMENT(error: string, conversationHistory: LLMMessage[]): string;
    /**
     * Generate validation feedback prompt
     *
     * @param validationResult Result from POWL validation
     * @param modelCode The generated model code
     * @returns Feedback prompt
     */
    VALIDATION_FEEDBACK(validationResult: {
        isValid: boolean;
        errors: string[];
        warnings: string[];
    }): string;
    /**
     * Extract model code from LLM response
     *
     * @param response The raw LLM response text
     * @returns Extracted Python code or null if not found
     */
    extractModelCode(response: string): string | null;
};
/**
 * Process model generation request
 */
export interface ModelGenerationRequest {
    processDescription: string;
    conversationHistory?: LLMMessage[];
}
/**
 * Process model generation response
 */
export interface ModelGenerationResponse {
    modelCode: string | null;
    rawResponse: string;
    conversationHistory: LLMMessage[];
}
/**
 * Main interface for LLM-based POWL generation
 */
export declare class LLMPowLGenerator {
    private conversationHistory;
    constructor();
    /**
     * Generate a prompt for the given process description
     */
    generatePrompt(processDescription: string): string;
    /**
     * Handle an error and generate refinement prompt
     */
    handleError(error: string): string;
    /**
     * Process LLM response and extract model code
     *
     * Note: This assumes the user prompt was already added to history via generatePrompt()
     * or was sent separately. If you need to track the user prompt, call generatePrompt()
     * and manually add it to history before calling this method.
     */
    processResponse(llmResponse: string): ModelGenerationResponse;
    /**
     * Reset conversation history (start fresh)
     */
    reset(): void;
    /**
     * Get current conversation history
     */
    getHistory(): LLMMessage[];
    /**
     * Add a user message to conversation history
     *
     * This is typically used after generating a prompt to track what was sent to the LLM
     *
     * @param content The user message content (typically the prompt)
     */
    addUserMessage(content: string): void;
    /**
     * Add an assistant message to conversation history
     *
     * This is typically used after receiving an LLM response
     *
     * @param content The assistant message content (typically the LLM response)
     */
    addAssistantMessage(content: string): void;
    /**
     * Generate validation feedback message
     */
    generateValidationFeedback(validationResult: {
        isValid: boolean;
        errors: string[];
        warnings: string[];
    }): string;
}
//# sourceMappingURL=llm-prompts.d.ts.map