zclaw_openfang/desktop/src/lib/cold-start-mapper.ts

/**
 * cold-start-mapper - Extract configuration from conversation content
 *
 * Maps user messages to cold start config (industry, name, personality, skills).
 * Uses keyword matching for deterministic extraction; LLM can refine later.
 */

// cold-start-mapper: keyword-based extraction for cold start configuration
// Future: LLM-based extraction fallback will use structured logger

// === Industry Detection ===

interface IndustryPattern {
  id: string;
  keywords: string[];
}

const INDUSTRY_PATTERNS: IndustryPattern[] = [
  {
    id: 'healthcare',
    keywords: ['医院', '医疗', '护士', '医生', '科室', '排班', '病历', '门诊', '住院', '行政', '护理', '医保', '挂号'],
  },
  {
    id: 'education',
    keywords: ['学校', '教育', '教师', '老师', '学生', '课程', '培训', '教学', '考试', '成绩', '教务', '班级'],
  },
  {
    id: 'garment',
    keywords: ['制衣', '服装', '面料', '打版', '缝纫', '裁床', '纺织', '生产', '工厂', '订单', '出货'],
  },
  {
    id: 'ecommerce',
    keywords: ['电商', '店铺', '商品', '库存', '物流', '客服', '促销', '直播', '选品', 'SKU', '运营', '零售'],
  },
];

export interface ColdStartMapping {
  detectedIndustry?: string;
  confidence: number;
  suggestedName?: string;
  personality?: { tone: string; formality: string; proactiveness: string };
  prioritySkills?: string[];
}

const INDUSTRY_SKILL_MAP: Record<string, string[]> = {
  healthcare: ['data_report', 'schedule_query', 'policy_search', 'meeting_notes'],
  education: ['data_report', 'schedule_query', 'content_writing', 'meeting_notes'],
  garment: ['data_report', 'schedule_query', 'inventory_mgmt', 'order_tracking'],
  ecommerce: ['data_report', 'inventory_mgmt', 'order_tracking', 'content_writing'],
};

const INDUSTRY_NAME_SUGGESTIONS: Record<string, string[]> = {
  healthcare: ['小医', '医管家', '康康'],
  education: ['小教', '学伴', '知了'],
  garment: ['小织', '裁缝', '布管家'],
  ecommerce: ['小商', '掌柜', '店小二'],
};

const INDUSTRY_PERSONALITY: Record<string, { tone: string; formality: string; proactiveness: string }> = {
  healthcare: { tone: 'professional', formality: 'formal', proactiveness: 'moderate' },
  education: { tone: 'friendly', formality: 'semi-formal', proactiveness: 'moderate' },
  garment: { tone: 'practical', formality: 'semi-formal', proactiveness: 'low' },
  ecommerce: { tone: 'energetic', formality: 'casual', proactiveness: 'high' },
};

/**
 * Detect industry from user message using keyword matching.
 */
export function detectIndustry(message: string): ColdStartMapping {
  if (!message || message.trim().length === 0) {
    return { confidence: 0 };
  }

  const lower = message.toLowerCase();
  let bestMatch = '';
  let bestScore = 0;

  for (const pattern of INDUSTRY_PATTERNS) {
    let score = 0;
    for (const keyword of pattern.keywords) {
      if (lower.includes(keyword)) {
        score += 1;
      }
    }
    if (score > bestScore) {
      bestScore = score;
      bestMatch = pattern.id;
    }
  }

  // Require at least 1 keyword match
  if (bestScore === 0) {
    return { confidence: 0 };
  }

  const confidence = Math.min(bestScore / 3, 1);

  const names = INDUSTRY_NAME_SUGGESTIONS[bestMatch] ?? [];
  const suggestedName = names.length > 0 ? names[0] : undefined;

  return {
    detectedIndustry: bestMatch,
    confidence,
    suggestedName,
    personality: INDUSTRY_PERSONALITY[bestMatch],
    prioritySkills: INDUSTRY_SKILL_MAP[bestMatch],
  };
}

/**
 * Detect if user is agreeing/confirming something.
 */
export function detectAffirmative(message: string): boolean {
  if (!message) return false;
  const affirmativePatterns = ['好', '可以', '行', '没问题', '是的', '对', '嗯', 'OK', 'ok', '确认', '同意'];
  const lower = message.toLowerCase().trim();
  return affirmativePatterns.some((p) => lower === p || lower.startsWith(p));
}

/**
 * Detect if user is rejecting something.
 */
export function detectNegative(message: string): boolean {
  if (!message) return false;
  const negativePatterns = ['不', '不要', '算了', '换一个', '换', '不好', '不行', '其他', '别的'];
  const lower = message.toLowerCase().trim();
  return negativePatterns.some((p) => lower === p || lower.startsWith(p));
}

/**
 * Detect if user provides a name suggestion.
 */
export function detectNameSuggestion(message: string): string | undefined {
  if (!message) return undefined;
  // Match patterns like "叫我小王" "叫XX" "用XX" "叫 XX 吧"
  const patterns = [/叫[我它他她]?[""''「」]?(\S{1,8})[""''「」]?[吧。！]?$/, /用[""''「」]?(\S{1,8})[""''「」]?[吧。！]?$/];
  for (const pattern of patterns) {
    const match = message.match(pattern);
    if (match && match[1]) {
      const name = match[1].replace(/[吧。！，、]/g, '').trim();
      if (name.length >= 1 && name.length <= 8) {
        return name;
      }
    }
  }
  return undefined;
}

/**
 * Determine the next cold start phase based on current phase and user message.
 */
export function determinePhaseTransition(
  currentPhase: string,
  userMessage: string,
): { nextPhase: string; mapping?: ColdStartMapping } | null {
  switch (currentPhase) {
    case 'agent_greeting': {
      const mapping = detectIndustry(userMessage);
      if (mapping.detectedIndustry && mapping.confidence > 0.3) {
        return { nextPhase: 'industry_discovery', mapping };
      }
      // User responded but no industry detected — keep probing
      return null;
    }

    case 'industry_discovery': {
      if (detectAffirmative(userMessage)) {
        return { nextPhase: 'identity_setup' };
      }
      if (detectNegative(userMessage)) {
        // Try to re-detect from the rejection
        const mapping = detectIndustry(userMessage);
        if (mapping.detectedIndustry) {
          return { nextPhase: 'industry_discovery', mapping };
        }
        return null;
      }
      // Direct industry mention
      const mapping = detectIndustry(userMessage);
      if (mapping.detectedIndustry) {
        return { nextPhase: 'identity_setup', mapping };
      }
      return null;
    }

    case 'identity_setup': {
      const customName = detectNameSuggestion(userMessage);
      if (customName) {
        return {
          nextPhase: 'first_task',
          mapping: { confidence: 1, suggestedName: customName },
        };
      }
      if (detectAffirmative(userMessage)) {
        return { nextPhase: 'first_task' };
      }
      if (detectNegative(userMessage)) {
        return null; // Stay in identity_setup for another suggestion
      }
      // User said something else, treat as name preference
      return {
        nextPhase: 'first_task',
        mapping: { confidence: 0.5, suggestedName: userMessage.trim().slice(0, 8) },
      };
    }

    case 'first_task': {
      // Any message in first_task is a real task — mark completed
      return { nextPhase: 'completed' };
    }

    default:
      return null;
  }
}