MOPC-Portal/src/server/services/ai-assignment.ts

/**
 * AI-Powered Assignment Service
 *
 * Uses GPT to analyze juror expertise and project requirements
 * to generate optimal assignment suggestions.
 *
 * Optimization:
 * - Batched processing (15 projects per batch)
 * - Description truncation (300 chars)
 * - Token tracking and cost logging
 *
 * GDPR Compliance:
 * - All data anonymized before AI processing
 * - IDs replaced with sequential identifiers
 * - No personal information sent to OpenAI
 */

import { getOpenAI, getConfiguredModel, buildCompletionParams } from '@/lib/openai'
import { logAIUsage, extractTokenUsage } from '@/server/utils/ai-usage'
import { classifyAIError, createParseError, logAIError } from './ai-errors'
import {
  anonymizeForAI,
  deanonymizeResults,
  validateAnonymization,
  DESCRIPTION_LIMITS,
  truncateAndSanitize,
  type AnonymizationResult,
} from './anonymization'

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

const ASSIGNMENT_BATCH_SIZE = 15

// Structured system prompt for assignment
const ASSIGNMENT_SYSTEM_PROMPT = `You are an expert jury assignment optimizer for an ocean conservation competition.

## Your Role
Match jurors to projects based on expertise alignment, workload balance, geographic diversity, and coverage requirements. You have access to rich data about both jurors and projects — use ALL available information to make optimal assignments.

## Available Data
- **Jurors**: expertiseTags (areas of expertise), bio (background description with deeper domain knowledge), country, currentAssignmentCount, maxAssignments
- **Projects**: title, description (detailed project overview), tags (with confidence 0-1), category (e.g. STARTUP, BUSINESS_CONCEPT), oceanIssue (focus area like CORAL_REEFS, POLLUTION), country, institution, teamSize, fileTypes (submitted document types)

## Matching Criteria (Weighted)
- Expertise & Domain Match (50%): How well juror tags, bio, and background align with project topics, category, ocean issue, and description. Use bio text to identify deeper domain expertise beyond explicit tags — e.g., a bio mentioning "20 years of coral research" matches coral-related projects even without explicit tags. Weight higher-confidence tags more heavily.
- Workload Balance (30%): Distribute assignments as evenly as possible; strongly prefer jurors below capacity. Never let one juror get significantly more assignments than another.
- Minimum Target (20%): Prioritize jurors who haven't reached their minimum assignment count

## Output Format
Return a JSON object:
{
  "assignments": [
    {
      "juror_id": "JUROR_001",
      "project_id": "PROJECT_001",
      "confidence_score": 0.0-1.0,
      "expertise_match_score": 0.0-1.0,
      "reasoning": "1-2 sentence justification referencing specific expertise matches"
    }
  ]
}

## Guidelines
- Each project MUST receive the required number of reviews — ensure full coverage
- Distribute assignments as evenly as possible across all jurors
- Do not assign jurors who are at or above their capacity
- Favor geographic diversity: avoid assigning jurors from the same country as the project when possible
- Consider disciplinary diversity: mix different expertise backgrounds per project
- confidence_score reflects overall assignment quality; expertise_match_score reflects tag/expertise overlap
- A strong match: shared expertise tags + relevant bio background + available capacity
- An acceptable match: related domain/ocean issue + available capacity
- A poor match: no expertise overlap, only assigned for coverage`

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

export interface AIAssignmentSuggestion {
  jurorId: string
  projectId: string
  confidenceScore: number // 0-1
  reasoning: string
  expertiseMatchScore: number // 0-1
}

export interface AIAssignmentResult {
  success: boolean
  suggestions: AIAssignmentSuggestion[]
  error?: string
  tokensUsed?: number
  fallbackUsed?: boolean
}

interface JurorForAssignment {
  id: string
  name?: string | null
  email: string
  expertiseTags: string[]
  bio?: string | null
  country?: string | null
  maxAssignments?: number | null
  _count?: {
    assignments: number
  }
}

interface ProjectForAssignment {
  id: string
  title: string
  description?: string | null
  tags: string[]
  tagConfidences?: Array<{ name: string; confidence: number }>
  teamName?: string | null
  competitionCategory?: string | null
  oceanIssue?: string | null
  country?: string | null
  institution?: string | null
  teamSize?: number
  fileTypes?: string[]
  _count?: {
    assignments: number
  }
}

interface AssignmentConstraints {
  requiredReviewsPerProject: number
  minAssignmentsPerJuror?: number
  maxAssignmentsPerJuror?: number
  jurorLimits?: Record<string, number> // userId -> personal max assignments
  existingAssignments: Array<{
    jurorId: string
    projectId: string
  }>
}

export interface AssignmentProgressCallback {
  (progress: {
    currentBatch: number
    totalBatches: number
    processedCount: number
    totalProjects: number
  }): Promise<void>
}

// ─── AI Processing ───────────────────────────────────────────────────────────

/**
 * Process a batch of projects for assignment suggestions
 */
async function processAssignmentBatch(
  openai: NonNullable<Awaited<ReturnType<typeof getOpenAI>>>,
  model: string,
  anonymizedData: AnonymizationResult,
  batchProjects: typeof anonymizedData.projects,
  batchMappings: typeof anonymizedData.projectMappings,
  constraints: AssignmentConstraints,
  userId?: string,
  entityId?: string
): Promise<{
  suggestions: AIAssignmentSuggestion[]
  tokensUsed: number
}> {
  const suggestions: AIAssignmentSuggestion[] = []
  let tokensUsed = 0

  // Build prompt with batch-specific data
  const userPrompt = buildBatchPrompt(
    anonymizedData.jurors,
    batchProjects,
    constraints,
    anonymizedData.jurorMappings,
    batchMappings
  )

  const MAX_PARSE_RETRIES = 2
  let parseAttempts = 0
  let response: Awaited<ReturnType<typeof openai.chat.completions.create>>

  try {
    const params = buildCompletionParams(model, {
      messages: [
        { role: 'system', content: ASSIGNMENT_SYSTEM_PROMPT },
        { role: 'user', content: userPrompt },
      ],
      jsonMode: true,
      temperature: 0.1,
      maxTokens: 4000,
    })

    try {
      response = await openai.chat.completions.create(params)
    } catch (apiError) {
      // Provide clearer error for model-related issues
      const errorMsg = apiError instanceof Error ? apiError.message : String(apiError)
      if (errorMsg.includes('model') || errorMsg.includes('does not exist')) {
        throw new Error(`Invalid AI model "${model}". Please check the model name in Settings > AI Configuration.`)
      }
      throw apiError
    }

    const usage = extractTokenUsage(response)
    tokensUsed = usage.totalTokens

    // Log batch usage
    await logAIUsage({
      userId,
      action: 'ASSIGNMENT',
      entityType: 'Round',
      entityId,
      model,
      promptTokens: usage.promptTokens,
      completionTokens: usage.completionTokens,
      totalTokens: usage.totalTokens,
      batchSize: batchProjects.length,
      itemsProcessed: batchProjects.length,
      status: 'SUCCESS',
    })

    // Parse with retry logic
    let parsed: {
      assignments: Array<{
        juror_id: string
        project_id: string
        confidence_score: number
        expertise_match_score: number
        reasoning: string
      }>
    }

    while (true) {
      try {
        const content = response.choices[0]?.message?.content
        if (!content) {
          // Check if response indicates an issue
          const finishReason = response.choices[0]?.finish_reason
          if (finishReason === 'content_filter') {
            throw new Error('AI response was filtered. Try a different model or simplify the project descriptions.')
          }
          if (!response.choices || response.choices.length === 0) {
            throw new Error(`No response from model "${model}". This model may not exist or may not be available. Please verify the model name.`)
          }
          throw new Error(`Empty response from AI model "${model}". The model may not support this type of request.`)
        }
        parsed = JSON.parse(content)
        break
      } catch (parseError) {
        if (parseError instanceof SyntaxError && parseAttempts < MAX_PARSE_RETRIES) {
          parseAttempts++
          console.warn(`[AI Assignment] JSON parse failed, retrying (${parseAttempts}/${MAX_PARSE_RETRIES})`)

          // Retry the API call with hint
          const retryParams = buildCompletionParams(model, {
            messages: [
              { role: 'system', content: ASSIGNMENT_SYSTEM_PROMPT },
              { role: 'user', content: userPrompt + '\n\nIMPORTANT: Please ensure valid JSON output.' },
            ],
            jsonMode: true,
            temperature: 0.1,
            maxTokens: 4000,
          })
          response = await openai.chat.completions.create(retryParams)
          const retryUsage = extractTokenUsage(response)
          tokensUsed += retryUsage.totalTokens
          continue
        }
        throw parseError
      }
    }

    // De-anonymize and add to suggestions
    const deanonymized = deanonymizeResults(
      (parsed.assignments || []).map((a) => ({
        jurorId: a.juror_id,
        projectId: a.project_id,
        confidenceScore: Math.min(1, Math.max(0, a.confidence_score)),
        expertiseMatchScore: Math.min(1, Math.max(0, a.expertise_match_score)),
        reasoning: a.reasoning,
      })),
      anonymizedData.jurorMappings,
      batchMappings
    )

    for (const item of deanonymized) {
      suggestions.push({
        jurorId: item.realJurorId,
        projectId: item.realProjectId,
        confidenceScore: item.confidenceScore,
        reasoning: item.reasoning,
        expertiseMatchScore: item.expertiseMatchScore,
      })
    }

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

      await logAIUsage({
        userId,
        action: 'ASSIGNMENT',
        entityType: 'Round',
        entityId,
        model,
        promptTokens: 0,
        completionTokens: 0,
        totalTokens: tokensUsed,
        batchSize: batchProjects.length,
        itemsProcessed: 0,
        status: 'ERROR',
        errorMessage: parseError.message,
      })
    } else {
      throw error
    }
  }

  return { suggestions, tokensUsed }
}

/**
 * Build prompt for a batch of projects
 */
function buildBatchPrompt(
  jurors: AnonymizationResult['jurors'],
  projects: AnonymizationResult['projects'],
  constraints: AssignmentConstraints,
  jurorMappings: AnonymizationResult['jurorMappings'],
  projectMappings: AnonymizationResult['projectMappings']
): string {
  // Map existing assignments to anonymous IDs
  const jurorIdMap = new Map(jurorMappings.map((m) => [m.realId, m.anonymousId]))
  const projectIdMap = new Map(projectMappings.map((m) => [m.realId, m.anonymousId]))

  const anonymousExisting = constraints.existingAssignments
    .map((a) => ({
      jurorId: jurorIdMap.get(a.jurorId),
      projectId: projectIdMap.get(a.projectId),
    }))
    .filter((a) => a.jurorId && a.projectId)

  // Build per-juror limits mapped to anonymous IDs
  let jurorLimitsStr = ''
  if (constraints.jurorLimits && Object.keys(constraints.jurorLimits).length > 0) {
    const anonymousLimits: Record<string, number> = {}
    for (const [realId, limit] of Object.entries(constraints.jurorLimits)) {
      const anonId = jurorIdMap.get(realId)
      if (anonId) {
        anonymousLimits[anonId] = limit
      }
    }
    if (Object.keys(anonymousLimits).length > 0) {
      jurorLimitsStr = `\nJUROR_LIMITS: ${JSON.stringify(anonymousLimits)} (per-juror max assignments, override global max)`
    }
  }

  return `JURORS: ${JSON.stringify(jurors)}
PROJECTS: ${JSON.stringify(projects)}
CONSTRAINTS: ${constraints.requiredReviewsPerProject} reviews/project, max ${constraints.maxAssignmentsPerJuror || 'unlimited'}/juror${jurorLimitsStr}
EXISTING: ${JSON.stringify(anonymousExisting)}
Return JSON: {"assignments": [...]}`
}

/**
 * Generate AI-powered assignment suggestions with batching
 */
export async function generateAIAssignments(
  jurors: JurorForAssignment[],
  projects: ProjectForAssignment[],
  constraints: AssignmentConstraints,
  userId?: string,
  entityId?: string,
  onProgress?: AssignmentProgressCallback
): Promise<AIAssignmentResult> {
  // Truncate descriptions before anonymization
  const truncatedProjects = projects.map((p) => ({
    ...p,
    description: truncateAndSanitize(p.description, DESCRIPTION_LIMITS.ASSIGNMENT),
  }))

  // Anonymize data before sending to AI
  const anonymizedData = anonymizeForAI(jurors, truncatedProjects)

  // Validate anonymization
  if (!validateAnonymization(anonymizedData)) {
    console.error('[AI Assignment] Anonymization validation failed, falling back to algorithm')
    return generateFallbackAssignments(jurors, projects, constraints)
  }

  try {
    const openai = await getOpenAI()

    if (!openai) {
      console.log('[AI Assignment] OpenAI not configured, using fallback algorithm')
      return generateFallbackAssignments(jurors, projects, constraints)
    }

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

    const allSuggestions: AIAssignmentSuggestion[] = []
    let totalTokens = 0

    // Process projects in batches
    const totalBatches = Math.ceil(anonymizedData.projects.length / ASSIGNMENT_BATCH_SIZE)

    for (let i = 0; i < anonymizedData.projects.length; i += ASSIGNMENT_BATCH_SIZE) {
      const batchProjects = anonymizedData.projects.slice(i, i + ASSIGNMENT_BATCH_SIZE)
      const batchMappings = anonymizedData.projectMappings.slice(i, i + ASSIGNMENT_BATCH_SIZE)
      const currentBatch = Math.floor(i / ASSIGNMENT_BATCH_SIZE) + 1

      console.log(`[AI Assignment] Processing batch ${currentBatch}/${totalBatches}`)

      const { suggestions, tokensUsed } = await processAssignmentBatch(
        openai,
        model,
        anonymizedData,
        batchProjects,
        batchMappings,
        constraints,
        userId,
        entityId
      )

      allSuggestions.push(...suggestions)
      totalTokens += tokensUsed

      // Report progress after each batch
      if (onProgress) {
        const processedCount = Math.min((currentBatch) * ASSIGNMENT_BATCH_SIZE, projects.length)
        await onProgress({
          currentBatch,
          totalBatches,
          processedCount,
          totalProjects: projects.length,
        })
      }
    }

    console.log(`[AI Assignment] Completed. Total suggestions: ${allSuggestions.length}, Total tokens: ${totalTokens}`)

    return {
      success: true,
      suggestions: allSuggestions,
      tokensUsed: totalTokens,
      fallbackUsed: false,
    }

  } catch (error) {
    const classified = classifyAIError(error)
    logAIError('Assignment', 'generateAIAssignments', classified)

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

    console.error('[AI Assignment] AI assignment failed, using fallback:', classified.message)
    return generateFallbackAssignments(jurors, projects, constraints)
  }
}

// ─── Fallback Algorithm ──────────────────────────────────────────────────────

/**
 * Fallback algorithm-based assignment when AI is unavailable
 */
export function generateFallbackAssignments(
  jurors: JurorForAssignment[],
  projects: ProjectForAssignment[],
  constraints: AssignmentConstraints
): AIAssignmentResult {
  const suggestions: AIAssignmentSuggestion[] = []
  const existingSet = new Set(
    constraints.existingAssignments.map((a) => `${a.jurorId}:${a.projectId}`)
  )

  // Track assignments per juror and project
  const jurorAssignments = new Map<string, number>()
  const projectAssignments = new Map<string, number>()

  // Initialize counts from existing assignments
  for (const assignment of constraints.existingAssignments) {
    jurorAssignments.set(
      assignment.jurorId,
      (jurorAssignments.get(assignment.jurorId) || 0) + 1
    )
    projectAssignments.set(
      assignment.projectId,
      (projectAssignments.get(assignment.projectId) || 0) + 1
    )
  }

  // Also include current assignment counts
  for (const juror of jurors) {
    const current = juror._count?.assignments || 0
    jurorAssignments.set(
      juror.id,
      Math.max(jurorAssignments.get(juror.id) || 0, current)
    )
  }

  for (const project of projects) {
    const current = project._count?.assignments || 0
    projectAssignments.set(
      project.id,
      Math.max(projectAssignments.get(project.id) || 0, current)
    )
  }

  // Sort projects by need (fewest assignments first)
  const sortedProjects = [...projects].sort((a, b) => {
    const aCount = projectAssignments.get(a.id) || 0
    const bCount = projectAssignments.get(b.id) || 0
    return aCount - bCount
  })

  // For each project, find best matching jurors
  for (const project of sortedProjects) {
    const currentProjectAssignments = projectAssignments.get(project.id) || 0
    const neededReviews = Math.max(
      0,
      constraints.requiredReviewsPerProject - currentProjectAssignments
    )

    if (neededReviews === 0) continue

    // Score all available jurors
    const scoredJurors = jurors
      .filter((juror) => {
        // Check not already assigned
        if (existingSet.has(`${juror.id}:${project.id}`)) return false

        // Check not at limit
        const currentAssignments = jurorAssignments.get(juror.id) || 0
        const maxAssignments =
          juror.maxAssignments ?? constraints.maxAssignmentsPerJuror ?? Infinity
        if (currentAssignments >= maxAssignments) return false

        return true
      })
      .map((juror) => {
        const currentLoad = jurorAssignments.get(juror.id) || 0
        const maxLoad = juror.maxAssignments ?? constraints.maxAssignmentsPerJuror ?? 20
        const minTarget = constraints.minAssignmentsPerJuror ?? 5

        return {
          juror,
          score: calculateExpertiseScore(juror.expertiseTags, project.tags, project.tagConfidences),
          loadScore: calculateLoadScore(currentLoad, maxLoad),
          underMinBonus: calculateUnderMinBonus(currentLoad, minTarget),
        }
      })
      .sort((a, b) => {
        // Combined score: 50% expertise, 30% load balancing, 20% under-min bonus
        const aTotal = a.score * 0.5 + a.loadScore * 0.3 + a.underMinBonus * 0.2
        const bTotal = b.score * 0.5 + b.loadScore * 0.3 + b.underMinBonus * 0.2
        return bTotal - aTotal
      })

    // Assign top jurors
    for (let i = 0; i < Math.min(neededReviews, scoredJurors.length); i++) {
      const { juror, score } = scoredJurors[i]

      suggestions.push({
        jurorId: juror.id,
        projectId: project.id,
        confidenceScore: score,
        expertiseMatchScore: score,
        reasoning: generateFallbackReasoning(
          juror.expertiseTags,
          project.tags,
          score
        ),
      })

      // Update tracking
      existingSet.add(`${juror.id}:${project.id}`)
      jurorAssignments.set(juror.id, (jurorAssignments.get(juror.id) || 0) + 1)
      projectAssignments.set(
        project.id,
        (projectAssignments.get(project.id) || 0) + 1
      )
    }
  }

  return {
    success: true,
    suggestions,
    fallbackUsed: true,
  }
}

/**
 * Calculate expertise match score based on tag overlap
 * When tagConfidences are available, weights matches by confidence
 */
function calculateExpertiseScore(
  jurorTags: string[],
  projectTags: string[],
  tagConfidences?: Array<{ name: string; confidence: number }>
): number {
  if (jurorTags.length === 0 || projectTags.length === 0) {
    return 0.5 // Neutral score if no tags
  }

  const jurorTagsLower = new Set(jurorTags.map((t) => t.toLowerCase()))

  // If we have confidence data, use weighted scoring
  if (tagConfidences && tagConfidences.length > 0) {
    let weightedMatches = 0
    let totalWeight = 0

    for (const tc of tagConfidences) {
      totalWeight += tc.confidence
      if (jurorTagsLower.has(tc.name.toLowerCase())) {
        weightedMatches += tc.confidence
      }
    }

    if (totalWeight === 0) return 0.5

    const weightedRatio = weightedMatches / totalWeight
    const hasExpertise = weightedMatches > 0 ? 0.2 : 0
    return Math.min(1, weightedRatio * 0.8 + hasExpertise)
  }

  // Fallback: unweighted matching using flat tags
  const matchingTags = projectTags.filter((t) =>
    jurorTagsLower.has(t.toLowerCase())
  )

  const matchRatio = matchingTags.length / projectTags.length
  const hasExpertise = matchingTags.length > 0 ? 0.2 : 0

  return Math.min(1, matchRatio * 0.8 + hasExpertise)
}

/**
 * Calculate load balancing score (higher score = less loaded)
 */
function calculateLoadScore(currentLoad: number, maxLoad: number): number {
  if (maxLoad === 0) return 0
  const utilization = currentLoad / maxLoad
  return Math.max(0, 1 - utilization)
}

/**
 * Calculate bonus for jurors under their minimum target
 * Returns 1.0 if under min, scaled down as approaching min
 */
function calculateUnderMinBonus(currentLoad: number, minTarget: number): number {
  if (currentLoad >= minTarget) return 0
  // Scale bonus based on how far under min (1.0 at 0 load, decreasing as approaching min)
  return (minTarget - currentLoad) / minTarget
}

/**
 * Generate reasoning for fallback assignments
 */
function generateFallbackReasoning(
  jurorTags: string[],
  projectTags: string[],
  score: number
): string {
  const jurorTagsLower = new Set(jurorTags.map((t) => t.toLowerCase()))
  const matchingTags = projectTags.filter((t) =>
    jurorTagsLower.has(t.toLowerCase())
  )

  if (matchingTags.length > 0) {
    return `Expertise match: ${matchingTags.join(', ')}. Match score: ${(score * 100).toFixed(0)}%.`
  }

  if (score >= 0.5) {
    return `Assigned for workload balance. No direct expertise match but available capacity.`
  }

  return `Assigned to ensure project coverage.`
}