Keep/keep-notes/lib/ai/services/auto-label-creation.service.ts

import { prisma } from '@/lib/prisma'
import { getAIProvider } from '@/lib/ai/factory'
import { getSystemConfig } from '@/lib/config'

export interface SuggestedLabel {
  name: string
  count: number
  confidence: number
  noteIds: string[]
}

export interface AutoLabelSuggestion {
  notebookId: string
  notebookName: string
  notebookIcon: string | null
  suggestedLabels: SuggestedLabel[]
  totalNotes: number
}

/**
 * Service for automatically suggesting new labels based on recurring themes
 * (Story 5.4 - IA4)
 */
export class AutoLabelCreationService {
  /**
   * Analyze a notebook and suggest new labels based on recurring themes
   * @param notebookId - Notebook ID to analyze
   * @param userId - User ID (for authorization)
   * @returns Suggested labels or null if not enough notes/no patterns found
   */
  async suggestLabels(notebookId: string, userId: string, language: string = 'en'): Promise<AutoLabelSuggestion | null> {
    // 1. Get notebook with existing labels
    const notebook = await prisma.notebook.findFirst({
      where: {
        id: notebookId,
        userId,
      },
      include: {
        labels: {
          select: {
            id: true,
            name: true,
          },
        },
        _count: {
          select: { notes: true },
        },
      },
    })

    if (!notebook) {
      throw new Error('Notebook not found')
    }

    // Only trigger if notebook has 15+ notes (PRD requirement)
    if (notebook._count.notes < 15) {
      return null
    }

    // Get all notes in this notebook
    const notes = await prisma.note.findMany({
      where: {
        notebookId,
        userId,
      },
      select: {
        id: true,
        title: true,
        content: true,
        labelRelations: {
          select: {
            name: true,
          },
        },
      },
      orderBy: {
        updatedAt: 'desc',
      },
      take: 100, // Limit to 100 most recent notes
    })

    if (notes.length === 0) {
      return null
    }

    // 2. Use AI to detect recurring themes
    const suggestions = await this.detectRecurringThemes(notes, notebook, language)

    return suggestions
  }

  /**
   * Use AI to detect recurring themes and suggest labels
   */
  private async detectRecurringThemes(
    notes: any[],
    notebook: any,
    language: string
  ): Promise<AutoLabelSuggestion | null> {
    const existingLabelNames = new Set<string>(
      notebook.labels.map((l: any) => l.name.toLowerCase())
    )

    const prompt = this.buildPrompt(notes, existingLabelNames, language)

    try {
      const config = await getSystemConfig()
      const provider = getAIProvider(config)
      const response = await provider.generateText(prompt)

      // Parse AI response
      const suggestions = this.parseAIResponse(response, notes)

      if (!suggestions || suggestions.suggestedLabels.length === 0) {
        return null
      }

      return {
        notebookId: notebook.id,
        notebookName: notebook.name,
        notebookIcon: notebook.icon,
        suggestedLabels: suggestions.suggestedLabels,
        totalNotes: notebook._count.notes,
      }
    } catch (error) {
      console.error('Failed to detect recurring themes:', error)
      return null
    }
  }

  /**
   * Build prompt for AI (localized)
   */
  private buildPrompt(notes: any[], existingLabelNames: Set<string>, language: string = 'en'): string {
    const notesSummary = notes
      .map((note, index) => {
        const title = note.title || 'Sans titre'
        const content = note.content.substring(0, 150)
        return `[${index}] "${title}": ${content}`
      })
      .join('\n')

    const existingLabels = Array.from(existingLabelNames).join(', ')

    const instructions: Record<string, string> = {
      fr: `
Tu es un assistant qui détecte les thèmes récurrents dans des notes pour suggérer de nouvelles étiquettes.

CARNET ANALYSÉ :
${notes.length} notes

ÉTIQUETTES EXISTANTES (ne pas suggérer celles-ci) :
${existingLabels || 'Aucune'}

NOTES DU CARNET :
${notesSummary}

TÂCHE :
Analyse les notes et détecte les thèmes récurrents (mots-clés, sujets, lieux, personnes).
Un thème doit apparaître dans au moins 5 notes différentes pour être suggéré.

FORMAT DE RÉPONSE (JSON) :
{
  "labels": [
    {
      "nom": "nom_du_label",
      "note_indices": [0, 5, 12, 23, 45],
      "confiance": 0.85
    }
  ]
}

RÈGLES :
- Le nom du label doit être court (1-2 mots max)
- Un thème doit apparaître dans 5+ notes pour être suggéré
- La confiance doit être > 0.60
- Ne pas suggérer des étiquettes qui existent déjà
- Priorise les lieux, personnes, catégories claires
- Maximum 5 suggestions

Exemples de bonnes étiquettes :
- "tokyo", "kyoto", "osaka" (lieux)
- "hôtels", "restos", "vols" (catégories)
- "marie", "jean", "équipe" (personnes)

Ta réponse (JSON seulement) :
`.trim(),
      en: `
You are an assistant that detects recurring themes in notes to suggest new labels.

ANALYZED NOTEBOOK:
${notes.length} notes

EXISTING LABELS (do not suggest these):
${existingLabels || 'None'}

NOTEBOOK NOTES:
${notesSummary}

TASK:
Analyze the notes and detect recurring themes (keywords, subjects, places, people).
A theme must appear in at least 5 different notes to be suggested.

RESPONSE FORMAT (JSON):
{
  "labels": [
    {
      "nom": "label_name",
      "note_indices": [0, 5, 12, 23, 45],
      "confiance": 0.85
    }
  ]
}

RULES:
- Label name must be short (max 1-2 words)
- A theme must appear in 5+ notes to be suggested
- Confidence must be > 0.60
- Do not suggest labels that already exist
- Prioritize places, people, clear categories
- Maximum 5 suggestions

Examples of good labels:
- "tokyo", "kyoto", "osaka" (places)
- "hotels", "restaurants", "flights" (categories)
- "mary", "john", "team" (people)

Your response (JSON only):
`.trim(),
      fa: `
شما یک دستیار هستید که تم‌های تکرارشونده در یادداشت‌ها را برای پیشنهاد برچسب‌های جدید شناسایی می‌کنید.

دفترچه‌ تحلیل شده:
${notes.length} یادداشت

برچسب‌های موجود (این‌ها را پیشنهاد ندهید):
${existingLabels || 'هیچ'}

یادداشت‌های دفترچه:
${notesSummary}

وظیفه:
یادداشت‌ها را تحلیل کنید و تم‌های تکرارشونده (کلمات کلیدی، موضوعات، مکان‌ها، افراد) را شناسایی کنید.
یک تم باید حداقل در ۵ یادداشت مختلف ظاهر شود تا پیشنهاد داده شود.

فرمت پاسخ (JSON):
{
  "labels": [
    {
      "nom": "نام_برچسب",
      "note_indices": [0, 5, 12, 23, 45],
      "confiance": 0.85
    }
  ]
}

قوانین:
- نام برچسب باید کوتاه باشد (حداکثر ۱-۲ کلمه)
- یک تم باید در ۵+ یادداشت ظاهر شود تا پیشنهاد داده شود
- اطمینان باید > 0.60 باشد
- برچسب‌هایی که قبلاً وجود دارند را پیشنهاد ندهید
- اولویت با مکان‌ها، افراد، دسته‌بندی‌های واضح است
- حداکثر ۵ پیشنهاد

مثال‌های برچسب خوب:
- "توکیو"، "کیوتو"، "اوزاکا" (مکان‌ها)
- "هتل‌ها"، "رستوران‌ها"، "پروازها" (دسته‌بندی‌ها)
- "مریم"، "علی"، "تیم" (افراد)

پاسخ شما (فقط JSON):
`.trim(),
      es: `
Eres un asistente que detecta temas recurrentes en notas para sugerir nuevas etiquetas.

CUADERNO ANALIZADO:
${notes.length} notas

ETIQUETAS EXISTENTES (no sugerir estas):
${existingLabels || 'Ninguna'}

NOTAS DEL CUADERNO:
${notesSummary}

TAREA:
Analiza las notas y detecta temas recurrentes (palabras clave, temas, lugares, personas).
Un tema debe aparecer en al menos 5 notas diferentes para ser sugerido.

FORMATO DE RESPUESTA (JSON):
{
  "labels": [
    {
      "nom": "nombre_etiqueta",
      "note_indices": [0, 5, 12, 23, 45],
      "confiance": 0.85
    }
  ]
}

REGLAS:
- El nombre de la etiqueta debe ser corto (máx 1-2 palabras)
- Un tema debe aparecer en 5+ notas para ser sugerido
- La confianza debe ser > 0.60
- No sugieras etiquetas que ya existen
- Prioriza lugares, personas, categorías claras
- Máximo 5 sugerencias

Ejemplos de buenas etiquetas:
- "tokio", "kyoto", "osaka" (lugares)
- "hoteles", "restaurantes", "vuelos" (categorías)
- "maría", "juan", "equipo" (personas)

Tu respuesta (solo JSON):
`.trim(),
      de: `
Du bist ein Assistent, der wiederkehrende Themen in Notizen erkennt, um neue Labels vorzuschlagen.

ANALYSIERTES NOTIZBUCH:
${notes.length} Notizen

VORHANDENE LABELS (schlage diese nicht vor):
${existingLabels || 'Keine'}

NOTIZBUCH-NOTIZEN:
${notesSummary}

AUFGABE:
Analysiere die Notizen und erkenne wiederkehrende Themen (Schlüsselwörter, Themen, Orte, Personen).
Ein Thema muss in mindestens 5 verschiedenen Notizen erscheinen, um vorgeschlagen zu werden.

ANTWORTFORMAT (JSON):
{
  "labels": [
    {
      "nom": "label_name",
      "note_indices": [0, 5, 12, 23, 45],
      "confiance": 0.85
    }
  ]
}

REGELN:
- Der Labelname muss kurz sein (max 1-2 Wörter)
- Ein Thema muss in 5+ Notizen erscheinen, um vorgeschlagen zu werden
- Konfidenz muss > 0.60 sein
- Schlage keine Labels vor, die bereits existieren
- Priorisiere Orte, Personen, klare Kategorien
- Maximal 5 Vorschläge

Beispiele für gute Labels:
- "tokio", "kyoto", "osaka" (Orte)
- "hotels", "restaurants", "flüge" (Kategorien)
- "maria", "johannes", "team" (Personen)

Deine Antwort (nur JSON):
`.trim()
    }

    return instructions[language] || instructions['en'] || instructions['fr']
  }

  /**
   * Parse AI response into suggested labels
   */
  private parseAIResponse(response: string, notes: any[]): { suggestedLabels: SuggestedLabel[] } | null {
    try {
      const jsonMatch = response.match(/\{[\s\S]*\}/)
      if (!jsonMatch) {
        throw new Error('No JSON found in response')
      }

      const aiData = JSON.parse(jsonMatch[0])

      const suggestedLabels: SuggestedLabel[] = (aiData.labels || [])
        .map((label: any) => {
          // Filter by confidence threshold
          if (label.confiance <= 0.60) return null

          // Get note IDs from indices
          const noteIds = label.note_indices
            .map((idx: number) => notes[idx]?.id)
            .filter(Boolean)

          // Must have at least 5 notes
          if (noteIds.length < 5) return null

          return {
            name: label.nom,
            count: noteIds.length,
            confidence: label.confiance,
            noteIds,
          }
        })
        .filter(Boolean)

      if (suggestedLabels.length === 0) {
        return null
      }

      // Sort by count (descending) and confidence
      suggestedLabels.sort((a, b) => {
        if (b.count !== a.count) {
          return b.count - a.count // More notes first
        }
        return b.confidence - a.confidence // Then higher confidence
      })

      // Limit to top 5
      return {
        suggestedLabels: suggestedLabels.slice(0, 5),
      }
    } catch (error) {
      console.error('Failed to parse AI response:', error)
      return null
    }
  }

  /**
   * Create suggested labels and assign them to notes
   * @param notebookId - Notebook ID
   * @param userId - User ID
   * @param suggestions - Suggested labels to create
   * @param selectedLabels - Labels user selected to create
   * @returns Number of labels created
   */
  async createLabels(
    notebookId: string,
    userId: string,
    suggestions: AutoLabelSuggestion,
    selectedLabels: string[]
  ): Promise<number> {
    let createdCount = 0

    for (const suggestedLabel of suggestions.suggestedLabels) {
      if (!selectedLabels.includes(suggestedLabel.name)) continue

      // Create the label
      const label = await prisma.label.create({
        data: {
          name: suggestedLabel.name,
          color: 'gray', // Default color, user can change later
          notebookId,
          userId,
        },
      })

      // Assign label to all suggested notes (updateMany doesn't support relations)
      for (const noteId of suggestedLabel.noteIds) {
        await prisma.note.update({
          where: { id: noteId },
          data: {
            labelRelations: {
              connect: {
                id: label.id,
              },
            },
          },
        })
      }

      createdCount++
    }

    return createdCount
  }
}

// Export singleton instance
export const autoLabelCreationService = new AutoLabelCreationService()