MOPC-Portal/src/server/services/mentor-matching.ts

/**
 * AI-Powered Mentor Matching Service
 *
 * Matches mentors to projects based on expertise alignment.
 *
 * Optimization:
 * - Batched processing (15 projects per batch)
 * - Token tracking and cost logging
 * - Fallback to algorithmic matching
 *
 * GDPR Compliance:
 * - All data anonymized before AI processing
 * - No personal information sent to OpenAI
 */

import { PrismaClient, OceanIssue, CompetitionCategory } from '@prisma/client'
import { getOpenAI, getConfiguredModel, buildCompletionParams } from '@/lib/openai'
import { logAIUsage, extractTokenUsage } from '@/server/utils/ai-usage'
import { classifyAIError, createParseError, logAIError } from './ai-errors'

// ─── Constants ───────────────────────────────────────────────────────────────

const MENTOR_BATCH_SIZE = 15

// Optimized system prompt
const MENTOR_MATCHING_SYSTEM_PROMPT = `Match mentors to projects by expertise. Return JSON.
Format for each project: {"matches": [{project_id, mentor_matches: [{mentor_index, confidence_score: 0-1, expertise_match_score: 0-1, reasoning: str}]}]}
Rank by suitability. Consider expertise alignment and availability.`

// ─── Types ───────────────────────────────────────────────────────────────────

interface ProjectInfo {
  id: string
  title: string
  description: string | null
  oceanIssue: OceanIssue | null
  competitionCategory: CompetitionCategory | null
  tags: string[]
}

interface MentorInfo {
  id: string
  name: string | null
  email: string
  expertiseTags: string[]
  currentAssignments: number
  maxAssignments: number | null
}

interface MentorMatch {
  mentorId: string
  confidenceScore: number
  expertiseMatchScore: number
  reasoning: string
}

// ─── Batched AI Matching ─────────────────────────────────────────────────────

/**
 * Process a batch of projects for mentor matching
 */
async function processMatchingBatch(
  openai: NonNullable<Awaited<ReturnType<typeof getOpenAI>>>,
  model: string,
  projects: ProjectInfo[],
  mentors: MentorInfo[],
  limit: number,
  userId?: string
): Promise<{
  results: Map<string, MentorMatch[]>
  tokensUsed: number
}> {
  const results = new Map<string, MentorMatch[]>()
  let tokensUsed = 0

  // Anonymize project data
  const anonymizedProjects = projects.map((p, index) => ({
    project_id: `P${index + 1}`,
    real_id: p.id,
    description: p.description?.slice(0, 350) || 'No description',
    category: p.competitionCategory,
    oceanIssue: p.oceanIssue,
    tags: p.tags,
  }))

  // Anonymize mentor data
  const anonymizedMentors = mentors.map((m, index) => ({
    index,
    expertise: m.expertiseTags,
    availability: m.maxAssignments
      ? `${m.currentAssignments}/${m.maxAssignments}`
      : 'unlimited',
  }))

  const userPrompt = `PROJECTS:
${anonymizedProjects.map(p => `${p.project_id}: Category=${p.category || 'N/A'}, Issue=${p.oceanIssue || 'N/A'}, Tags=[${p.tags.join(', ')}], Desc=${p.description.slice(0, 200)}`).join('\n')}

MENTORS:
${anonymizedMentors.map(m => `${m.index}: Expertise=[${m.expertise.join(', ')}], Availability=${m.availability}`).join('\n')}

For each project, rank top ${limit} mentors.`

  try {
    const params = buildCompletionParams(model, {
      messages: [
        { role: 'system', content: MENTOR_MATCHING_SYSTEM_PROMPT },
        { role: 'user', content: userPrompt },
      ],
      jsonMode: true,
      temperature: 0.3,
      maxTokens: 4000,
    })

    const response = await openai.chat.completions.create(params)
    const usage = extractTokenUsage(response)
    tokensUsed = usage.totalTokens

    // Log usage
    await logAIUsage({
      userId,
      action: 'MENTOR_MATCHING',
      entityType: 'Project',
      model,
      promptTokens: usage.promptTokens,
      completionTokens: usage.completionTokens,
      totalTokens: usage.totalTokens,
      batchSize: projects.length,
      itemsProcessed: projects.length,
      status: 'SUCCESS',
    })

    const content = response.choices[0]?.message?.content
    if (!content) {
      throw new Error('No response from AI')
    }

    const parsed = JSON.parse(content) as {
      matches: Array<{
        project_id: string
        mentor_matches: Array<{
          mentor_index: number
          confidence_score: number
          expertise_match_score: number
          reasoning: string
        }>
      }>
    }

    // Map results back to real IDs
    for (const projectMatch of parsed.matches || []) {
      const project = anonymizedProjects.find(p => p.project_id === projectMatch.project_id)
      if (!project) continue

      const mentorMatches: MentorMatch[] = []
      for (const match of projectMatch.mentor_matches || []) {
        if (match.mentor_index >= 0 && match.mentor_index < mentors.length) {
          mentorMatches.push({
            mentorId: mentors[match.mentor_index].id,
            confidenceScore: Math.min(1, Math.max(0, match.confidence_score)),
            expertiseMatchScore: Math.min(1, Math.max(0, match.expertise_match_score)),
            reasoning: match.reasoning,
          })
        }
      }
      results.set(project.real_id, mentorMatches)
    }

  } catch (error) {
    if (error instanceof SyntaxError) {
      const parseError = createParseError(error.message)
      logAIError('MentorMatching', 'batch processing', parseError)

      await logAIUsage({
        userId,
        action: 'MENTOR_MATCHING',
        entityType: 'Project',
        model,
        promptTokens: 0,
        completionTokens: 0,
        totalTokens: tokensUsed,
        batchSize: projects.length,
        itemsProcessed: 0,
        status: 'ERROR',
        errorMessage: parseError.message,
      })

      // Return empty results for batch (will fall back to algorithm)
      for (const project of projects) {
        results.set(project.id, [])
      }
    } else {
      throw error
    }
  }

  return { results, tokensUsed }
}

/**
 * Get AI-suggested mentor matches for multiple projects (batched)
 */
export async function getAIMentorSuggestionsBatch(
  prisma: PrismaClient,
  projectIds: string[],
  limit: number = 5,
  userId?: string
): Promise<Map<string, MentorMatch[]>> {
  const allResults = new Map<string, MentorMatch[]>()

  // Get projects
  const projects = await prisma.project.findMany({
    where: { id: { in: projectIds } },
    select: {
      id: true,
      title: true,
      description: true,
      oceanIssue: true,
      competitionCategory: true,
      tags: true,
    },
  })

  if (projects.length === 0) {
    return allResults
  }

  // Get available mentors
  const mentors = await prisma.user.findMany({
    where: {
      OR: [
        { expertiseTags: { isEmpty: false } },
        { role: 'JURY_MEMBER' },
      ],
      status: 'ACTIVE',
    },
    select: {
      id: true,
      name: true,
      email: true,
      expertiseTags: true,
      maxAssignments: true,
      mentorAssignments: {
        select: { id: true },
      },
    },
  })

  // Filter mentors who haven't reached max assignments
  const availableMentors: MentorInfo[] = mentors
    .filter((m) => {
      const currentAssignments = m.mentorAssignments.length
      return !m.maxAssignments || currentAssignments < m.maxAssignments
    })
    .map((m) => ({
      id: m.id,
      name: m.name,
      email: m.email,
      expertiseTags: m.expertiseTags,
      currentAssignments: m.mentorAssignments.length,
      maxAssignments: m.maxAssignments,
    }))

  if (availableMentors.length === 0) {
    return allResults
  }

  // Try AI matching
  try {
    const openai = await getOpenAI()
    if (!openai) {
      console.log('[Mentor Matching] OpenAI not configured, using algorithm')
      return getAlgorithmicMatchesBatch(projects, availableMentors, limit)
    }

    const model = await getConfiguredModel()
    console.log(`[Mentor Matching] Using model: ${model} for ${projects.length} projects in batches of ${MENTOR_BATCH_SIZE}`)

    let totalTokens = 0

    // Process in batches
    for (let i = 0; i < projects.length; i += MENTOR_BATCH_SIZE) {
      const batchProjects = projects.slice(i, i + MENTOR_BATCH_SIZE)

      console.log(`[Mentor Matching] Processing batch ${Math.floor(i / MENTOR_BATCH_SIZE) + 1}/${Math.ceil(projects.length / MENTOR_BATCH_SIZE)}`)

      const { results, tokensUsed } = await processMatchingBatch(
        openai,
        model,
        batchProjects,
        availableMentors,
        limit,
        userId
      )

      totalTokens += tokensUsed

      // Merge results
      for (const [projectId, matches] of results) {
        allResults.set(projectId, matches)
      }
    }

    console.log(`[Mentor Matching] Completed. Total tokens: ${totalTokens}`)

    // Fill in any missing projects with algorithmic fallback
    for (const project of projects) {
      if (!allResults.has(project.id) || allResults.get(project.id)?.length === 0) {
        const fallbackMatches = getAlgorithmicMatches(project, availableMentors, limit)
        allResults.set(project.id, fallbackMatches)
      }
    }

    return allResults

  } catch (error) {
    const classified = classifyAIError(error)
    logAIError('MentorMatching', 'getAIMentorSuggestionsBatch', classified)

    // Log failed attempt
    await logAIUsage({
      userId,
      action: 'MENTOR_MATCHING',
      entityType: 'Project',
      model: 'unknown',
      promptTokens: 0,
      completionTokens: 0,
      totalTokens: 0,
      batchSize: projects.length,
      itemsProcessed: 0,
      status: 'ERROR',
      errorMessage: classified.message,
    })

    console.error('[Mentor Matching] AI failed, using algorithm:', classified.message)
    return getAlgorithmicMatchesBatch(projects, availableMentors, limit)
  }
}

/**
 * Get AI-suggested mentor matches for a single project
 */
export async function getAIMentorSuggestions(
  prisma: PrismaClient,
  projectId: string,
  limit: number = 5,
  userId?: string
): Promise<MentorMatch[]> {
  const results = await getAIMentorSuggestionsBatch(prisma, [projectId], limit, userId)
  return results.get(projectId) || []
}

// ─── Algorithmic Fallback ────────────────────────────────────────────────────

/**
 * Algorithmic fallback for multiple projects
 */
function getAlgorithmicMatchesBatch(
  projects: ProjectInfo[],
  mentors: MentorInfo[],
  limit: number
): Map<string, MentorMatch[]> {
  const results = new Map<string, MentorMatch[]>()

  for (const project of projects) {
    results.set(project.id, getAlgorithmicMatches(project, mentors, limit))
  }

  return results
}

/**
 * Algorithmic fallback for mentor matching
 */
function getAlgorithmicMatches(
  project: ProjectInfo,
  mentors: MentorInfo[],
  limit: number
): MentorMatch[] {
  // Build keyword set from project
  const projectKeywords = new Set<string>()

  if (project.oceanIssue) {
    projectKeywords.add(project.oceanIssue.toLowerCase().replace(/_/g, ' '))
  }

  if (project.competitionCategory) {
    projectKeywords.add(project.competitionCategory.toLowerCase().replace(/_/g, ' '))
  }

  project.tags.forEach((tag) => {
    tag.toLowerCase().split(/\s+/).forEach((word) => {
      if (word.length > 3) projectKeywords.add(word)
    })
  })

  if (project.description) {
    const words = project.description.toLowerCase().split(/\s+/)
    words.forEach((word) => {
      if (word.length > 4) projectKeywords.add(word.replace(/[^a-z]/g, ''))
    })
  }

  // Score each mentor
  const scored = mentors.map((mentor) => {
    const mentorKeywords = new Set<string>()
    mentor.expertiseTags.forEach((tag) => {
      tag.toLowerCase().split(/\s+/).forEach((word) => {
        if (word.length > 2) mentorKeywords.add(word)
      })
    })

    // Calculate overlap
    let matchCount = 0
    projectKeywords.forEach((keyword) => {
      mentorKeywords.forEach((mentorKeyword) => {
        if (keyword.includes(mentorKeyword) || mentorKeyword.includes(keyword)) {
          matchCount++
        }
      })
    })

    const expertiseMatchScore = mentorKeywords.size > 0
      ? Math.min(1, matchCount / mentorKeywords.size)
      : 0

    // Factor in availability
    const availabilityScore = mentor.maxAssignments
      ? 1 - (mentor.currentAssignments / mentor.maxAssignments)
      : 1

    const confidenceScore = (expertiseMatchScore * 0.7 + availabilityScore * 0.3)

    return {
      mentorId: mentor.id,
      confidenceScore: Math.round(confidenceScore * 100) / 100,
      expertiseMatchScore: Math.round(expertiseMatchScore * 100) / 100,
      reasoning: `Matched ${matchCount} keyword(s). Availability: ${availabilityScore > 0.5 ? 'Good' : 'Limited'}.`,
    }
  })

  // Sort by confidence and return top matches
  return scored
    .sort((a, b) => b.confidenceScore - a.confidenceScore)
    .slice(0, limit)
}

/**
 * Round-robin assignment for load balancing
 */
export async function getRoundRobinMentor(
  prisma: PrismaClient,
  excludeMentorIds: string[] = []
): Promise<string | null> {
  const mentors = await prisma.user.findMany({
    where: {
      OR: [
        { expertiseTags: { isEmpty: false } },
        { role: 'JURY_MEMBER' },
      ],
      status: 'ACTIVE',
      id: { notIn: excludeMentorIds },
    },
    select: {
      id: true,
      maxAssignments: true,
      mentorAssignments: {
        select: { id: true },
      },
    },
    orderBy: {
      mentorAssignments: {
        _count: 'asc',
      },
    },
  })

  // Find mentor with fewest assignments who hasn't reached max
  for (const mentor of mentors) {
    const currentCount = mentor.mentorAssignments.length
    if (!mentor.maxAssignments || currentCount < mentor.maxAssignments) {
      return mentor.id
    }
  }

  return null
}