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

/**
 * AI Shortlist Service
 *
 * Generates ranked recommendations at end of evaluation rounds.
 * Follows patterns from ai-filtering.ts and ai-evaluation-summary.ts.
 *
 * GDPR Compliance:
 * - All project data is anonymized before AI processing
 * - No personal identifiers in prompts or responses
 */

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

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

export type ShortlistResult = {
  success: boolean
  recommendations: ShortlistRecommendation[]
  errors?: string[]
  tokensUsed?: number
}

export type ShortlistRecommendation = {
  projectId: string
  rank: number
  score: number
  strengths: string[]
  concerns: string[]
  recommendation: string
}

// ─── Main Function ──────────────────────────────────────────────────────────

/**
 * Generate an AI shortlist for projects in a round.
 * Only runs if EvaluationConfig.generateAiShortlist is true.
 */
export async function generateShortlist(
  params: {
    roundId: string
    competitionId: string
    category?: string
    topN?: number
    rubric?: string
  },
  prisma: PrismaClient | any,
): Promise<ShortlistResult> {
  const { roundId, competitionId, category, topN = 10, rubric } = params

  try {
    // Load projects with evaluations
    const where: Record<string, unknown> = {
      assignments: { some: { roundId } },
    }
    if (category) {
      where.competitionCategory = category
    }

    const projects = await prisma.project.findMany({
      where,
      include: {
        assignments: {
          where: { roundId },
          include: {
            evaluation: true,
          },
        },
        projectTags: { include: { tag: true } },
        files: { select: { id: true, type: true } },
        teamMembers: { select: { user: { select: { name: true } } } },
      },
    })

    if (projects.length === 0) {
      return {
        success: true,
        recommendations: [],
        errors: ['No projects found for this round'],
      }
    }

    // Aggregate scores per project
    const projectSummaries = projects.map((project: any) => {
      const evaluations = project.assignments
        .map((a: any) => a.evaluation)
        .filter(Boolean)
        .filter((e: any) => e.status === 'SUBMITTED')

      const scores = evaluations.map((e: any) => e.globalScore ?? 0)
      const avgScore = scores.length > 0
        ? scores.reduce((sum: number, s: number) => sum + s, 0) / scores.length
        : 0

      const feedbacks = evaluations
        .map((e: any) => e.feedbackGeneral)
        .filter(Boolean)

      return {
        id: project.id,
        title: project.title,
        description: project.description,
        category: project.competitionCategory,
        tags: project.projectTags.map((pt: any) => pt.tag.name),
        avgScore,
        evaluationCount: evaluations.length,
        feedbackSamples: feedbacks.slice(0, 3), // Max 3 feedback samples
      }
    })

    // Anonymize for AI
    const anonymized = projectSummaries.map((p: any, index: number) => ({
      anonymousId: `PROJECT_${String(index + 1).padStart(3, '0')}`,
      ...p,
      // Strip identifying info
      title: undefined,
      id: undefined,
    }))

    // Build idMap for de-anonymization
    const idMap = new Map<string, string>()
    projectSummaries.forEach((p: any, index: number) => {
      idMap.set(`PROJECT_${String(index + 1).padStart(3, '0')}`, p.id)
    })

    // Build prompt
    const systemPrompt = `You are a senior jury advisor for the Monaco Ocean Protection Challenge.

## Your Role
Analyze aggregated evaluation data to produce a ranked shortlist of top projects.

## Ranking Criteria (Weighted)
- Evaluation Scores (40%): Average scores across all jury evaluations
- Innovation & Impact (25%): Novelty of approach and potential environmental impact
- Feasibility (20%): Likelihood of successful implementation
- Alignment (15%): Fit with ocean protection mission and competition goals

## Output Format
Return a JSON array:
[
  {
    "anonymousId": "PROJECT_001",
    "rank": 1,
    "score": 0-100,
    "strengths": ["strength 1", "strength 2"],
    "concerns": ["concern 1"],
    "recommendation": "1-2 sentence recommendation",
    "criterionBreakdown": {
      "evaluationScores": 38,
      "innovationImpact": 22,
      "feasibility": 18,
      "alignment": 14
    }
  }
]

## Guidelines
- Only include the requested number of top projects
- Score should reflect weighted combination of all criteria
- Be specific in strengths and concerns — avoid generic statements
- Consider feedback themes and evaluator consensus
- Higher evaluator consensus should boost confidence in ranking`

    const userPrompt = `Analyze these anonymized project evaluations and produce a ranked shortlist of the top ${topN} projects.

${rubric ? `Evaluation rubric:\n${rubric}\n\n` : ''}Projects:
${JSON.stringify(anonymized, null, 2)}

Return a JSON array following the format specified in your instructions. Only include the top ${topN} projects. Rank by overall quality considering scores and feedback.`

    const openai = await getOpenAI()
    const model = await getConfiguredModel()

    if (!openai) {
      return {
        success: false,
        recommendations: [],
        errors: ['OpenAI client not configured'],
      }
    }

    const MAX_PARSE_RETRIES = 2
    let parseAttempts = 0
    let response = await openai.chat.completions.create(
      buildCompletionParams(model, {
        messages: [
          { role: 'system', content: systemPrompt },
          { role: 'user', content: userPrompt },
        ],
        temperature: 0.1,
        jsonMode: true,
      }),
    )

    let tokenUsage = extractTokenUsage(response)

    await logAIUsage({
      action: 'FILTERING',
      model,
      promptTokens: tokenUsage.promptTokens,
      completionTokens: tokenUsage.completionTokens,
      totalTokens: tokenUsage.totalTokens,
      status: 'SUCCESS',
    })

    // Parse response with retry logic
    let parsed: any[]
    while (true) {
      try {
        const content = response.choices[0]?.message?.content
        if (!content) {
          return {
            success: false,
            recommendations: [],
            errors: ['Empty AI response'],
            tokensUsed: tokenUsage.totalTokens,
          }
        }

        const json = JSON.parse(content)
        parsed = Array.isArray(json) ? json : json.rankings ?? json.projects ?? json.shortlist ?? []
        break
      } catch (parseError) {
        if (parseError instanceof SyntaxError && parseAttempts < MAX_PARSE_RETRIES) {
          parseAttempts++
          console.warn(`[AI Shortlist] JSON parse failed, retrying (${parseAttempts}/${MAX_PARSE_RETRIES})`)

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

        return {
          success: false,
          recommendations: [],
          errors: ['Failed to parse AI response as JSON'],
          tokensUsed: tokenUsage.totalTokens,
        }
      }
    }

    // De-anonymize and build recommendations
    const recommendations: ShortlistRecommendation[] = parsed
      .filter((item: any) => item.anonymousId && idMap.has(item.anonymousId))
      .map((item: any) => ({
        projectId: idMap.get(item.anonymousId)!,
        rank: item.rank ?? 0,
        score: item.score ?? 0,
        strengths: item.strengths ?? [],
        concerns: item.concerns ?? [],
        recommendation: item.recommendation ?? '',
      }))
      .sort((a: ShortlistRecommendation, b: ShortlistRecommendation) => a.rank - b.rank)

    return {
      success: true,
      recommendations,
      tokensUsed: tokenUsage.totalTokens,
    }
  } catch (error) {
    const classification = classifyAIError(error)
    logAIError('ai-shortlist', 'generateShortlist', classification)
    console.error('[AIShortlist] generateShortlist failed:', error)

    return {
      success: false,
      recommendations: [],
      errors: [error instanceof Error ? error.message : 'AI shortlist generation failed'],
    }
  }
}