goodgo-platform/apps/api/src/modules/analytics/infrastructure/services/avm-retrain-cron.service.ts

import { Injectable } from '@nestjs/common';
import { Cron } from '@nestjs/schedule';
import { PrismaService, LoggerService } from '@modules/shared';

@Injectable()
export class AvmRetrainCronService {
  private readonly aiServiceUrl: string;
  private readonly aiServiceApiKey: string;

  constructor(
    private readonly prisma: PrismaService,
    private readonly logger: LoggerService,
  ) {
    this.aiServiceUrl = process.env['AI_SERVICE_URL'] ?? 'http://localhost:8000';
    this.aiServiceApiKey = process.env['AI_SERVICE_API_KEY'] ?? '';
  }

  /**
   * Weekly retrain — every Sunday at 3 AM.
   *
   * 1. Export training data from database to the AI service
   * 2. Trigger ensemble retraining via POST /avm/v2/train
   * 3. Log results (version, metrics)
   */
  @Cron('0 3 * * 0', { name: 'avm-v2-weekly-retrain' })
  async weeklyRetrain(): Promise<void> {
    this.logger.log('Starting weekly AVM v2 retrain...', 'AvmRetrainCronService');

    try {
      // Step 1: Export training data
      const trainingData = await this.exportTrainingData();
      if (trainingData.length < 50) {
        this.logger.warn(
          `Insufficient training data (${trainingData.length} rows). Skipping retrain.`,
          'AvmRetrainCronService',
        );
        return;
      }

      // Step 2: Upload training data to AI service
      await this.uploadTrainingData(trainingData);

      // Step 3: Trigger retraining
      const result = await this.triggerRetrain();

      this.logger.log(
        `AVM v2 retrain completed: version=${result.model_version}, ` +
          `MAPE=${result.metrics?.mape ?? 'N/A'}%, ` +
          `samples=${result.training_samples}`,
        'AvmRetrainCronService',
      );
    } catch (err) {
      this.logger.error(
        `AVM v2 weekly retrain failed: ${(err as Error).message}`,
        undefined,
        'AvmRetrainCronService',
      );
    }
  }

  /**
   * Export property + listing + market data as training rows.
   *
   * Each row maps to the feature columns expected by the Python
   * AVM v2 training pipeline (see avm_v2_service._prepare_training_data).
   */
  async exportTrainingData(): Promise<TrainingRow[]> {
    const rows = await this.prisma.$queryRaw<RawTrainingRow[]>`
      WITH market AS (
        SELECT
          mi.district,
          mi.city,
          mi."avgPriceM2"          AS avg_price_m2,
          mi."totalListings"       AS listing_density,
          COALESCE(mi."absorptionRate", 0)  AS absorption_rate,
          mi."daysOnMarket"        AS dom_avg,
          COALESCE(mi."yoyChange", 0)       AS yoy_change
        FROM "MarketIndex" mi
        WHERE mi.period = (
          SELECT MAX(period) FROM "MarketIndex"
        )
      )
      SELECT
        p."propertyType"::text                              AS property_type,
        p."areaM2"                                          AS area_m2,
        COALESCE(p.bedrooms, 2)                             AS rooms,
        COALESCE(p.floor, 0)                                AS floor_level,
        COALESCE(p."totalFloors", p.floors, 0)              AS total_floors,
        COALESCE(p.direction::text, 'unknown')              AS direction,
        CASE
          WHEN p."totalFloors" > 0 AND p."areaM2" > 0
          THEN (p."totalFloors"::float * p."areaM2") / NULLIF(p."areaM2", 0)
          ELSE 1.0
        END                                                 AS floor_ratio,
        CASE
          WHEN p."yearBuilt" IS NOT NULL
          THEN EXTRACT(YEAR FROM NOW())::int - p."yearBuilt"
          ELSE 5
        END                                                 AS building_age_years,
        CASE WHEN p.amenities::text ILIKE '%elevator%' THEN 1.0 ELSE 0.0 END   AS has_elevator,
        CASE WHEN p.amenities::text ILIKE '%parking%'  THEN 1.0 ELSE 0.0 END   AS has_parking,
        CASE WHEN p.amenities::text ILIKE '%pool%'     THEN 1.0 ELSE 0.0 END   AS has_pool,
        CASE
          WHEN p."legalStatus" IN ('so_do', 'so_hong', 'SO_DO', 'SO_HONG') THEN 1.0
          ELSE 0.0
        END                                                 AS has_legal_paper,
        0.5                                                 AS developer_reputation,
        0.5                                                 AS neighborhood_score,
        COALESCE(
          ST_Distance(
            p.location::geography,
            ST_SetSRID(ST_MakePoint(106.6297, 10.8231), 4326)::geography
          ) / 1000.0,
          10.0
        )                                                   AS distance_to_cbd_km,
        COALESCE(p."metroDistanceM" / 1000.0, 5.0)         AS distance_to_metro_km,
        5.0                                                 AS distance_to_school_km,
        3.0                                                 AS distance_to_hospital_km,
        2.0                                                 AS distance_to_park_km,
        4.0                                                 AS distance_to_mall_km,
        0.1                                                 AS flood_zone_risk,
        COALESCE(m.avg_price_m2, 0)                         AS avg_price_district_3m_vnd_m2,
        COALESCE(m.listing_density, 0)                      AS listing_density,
        COALESCE(m.absorption_rate, 0)                      AS absorption_rate,
        COALESCE(m.dom_avg, 30)                             AS dom_avg,
        0.0                                                 AS price_momentum_30d,
        COALESCE(m.yoy_change, 0)                           AS yoy_change,
        0.5                                                 AS renovation_score,
        0.5                                                 AS view_quality,
        0.5                                                 AS interior_quality,
        0.3                                                 AS noise_level,
        0.5                                                 AS natural_light,
        EXTRACT(MONTH FROM l."publishedAt")::int            AS month,
        p.district                                          AS district,
        l."priceVND"::float                                 AS price_vnd
      FROM "Listing" l
      JOIN "Property" p ON l."propertyId" = p.id
      LEFT JOIN market m ON m.district = p.district AND m.city = p.city
      WHERE l.status IN ('ACTIVE', 'SOLD', 'RENTED')
        AND l."priceVND" > 100000000
        AND l."publishedAt" IS NOT NULL
        AND p."areaM2" > 0
      ORDER BY l."publishedAt" DESC
      LIMIT 50000
    `;

    return rows.map((r) => ({
      property_type: String(r.property_type).toLowerCase(),
      area_m2: Number(r.area_m2),
      rooms: Number(r.rooms),
      floor_level: Number(r.floor_level),
      total_floors: Number(r.total_floors),
      direction: String(r.direction).toLowerCase(),
      floor_ratio: Number(r.floor_ratio),
      building_age_years: Number(r.building_age_years),
      has_elevator: Number(r.has_elevator),
      has_parking: Number(r.has_parking),
      has_pool: Number(r.has_pool),
      has_legal_paper: Number(r.has_legal_paper),
      developer_reputation: Number(r.developer_reputation),
      neighborhood_score: Number(r.neighborhood_score),
      distance_to_cbd_km: Number(r.distance_to_cbd_km),
      distance_to_metro_km: Number(r.distance_to_metro_km),
      distance_to_school_km: Number(r.distance_to_school_km),
      distance_to_hospital_km: Number(r.distance_to_hospital_km),
      distance_to_park_km: Number(r.distance_to_park_km),
      distance_to_mall_km: Number(r.distance_to_mall_km),
      flood_zone_risk: Number(r.flood_zone_risk),
      avg_price_district_3m_vnd_m2: Number(r.avg_price_district_3m_vnd_m2),
      listing_density: Number(r.listing_density),
      absorption_rate: Number(r.absorption_rate),
      dom_avg: Number(r.dom_avg),
      price_momentum_30d: Number(r.price_momentum_30d),
      yoy_change: Number(r.yoy_change),
      renovation_score: Number(r.renovation_score),
      view_quality: Number(r.view_quality),
      interior_quality: Number(r.interior_quality),
      noise_level: Number(r.noise_level),
      natural_light: Number(r.natural_light),
      month: Number(r.month),
      district: String(r.district),
      price_vnd: Number(r.price_vnd),
    }));
  }

  private async uploadTrainingData(rows: TrainingRow[]): Promise<void> {
    const headers = Object.keys(rows[0]!);
    const csvLines = [headers.join(',')];
    for (const row of rows) {
      csvLines.push(headers.map((h) => String(row[h as keyof TrainingRow])).join(','));
    }
    const csv = csvLines.join('\n');

    const url = `${this.aiServiceUrl}/avm/v2/upload-training-data`;
    const reqHeaders: Record<string, string> = { 'Content-Type': 'text/csv' };
    if (this.aiServiceApiKey) {
      reqHeaders['X-API-Key'] = this.aiServiceApiKey;
    }

    const response = await fetch(url, {
      method: 'POST',
      headers: reqHeaders,
      body: csv,
      signal: AbortSignal.timeout(30_000),
    });

    if (!response.ok) {
      const text = await response.text().catch(() => '');
      throw new Error(`Training data upload failed (${response.status}): ${text}`);
    }

    this.logger.log(
      `Uploaded ${rows.length} training rows to AI service`,
      'AvmRetrainCronService',
    );
  }

  private async triggerRetrain(): Promise<RetrainResult> {
    const url = `${this.aiServiceUrl}/avm/v2/train`;
    const headers: Record<string, string> = { 'Content-Type': 'application/json' };
    if (this.aiServiceApiKey) {
      headers['X-API-Key'] = this.aiServiceApiKey;
    }

    const response = await fetch(url, {
      method: 'POST',
      headers,
      body: JSON.stringify({
        optuna_trials: 50,
        test_size: 0.15,
        val_size: 0.15,
      }),
      signal: AbortSignal.timeout(600_000), // 10 min — training can take a while
    });

    if (!response.ok) {
      const text = await response.text().catch(() => '');
      throw new Error(`Retrain request failed (${response.status}): ${text}`);
    }

    return response.json() as Promise<RetrainResult>;
  }
}

interface RawTrainingRow {
  property_type: string;
  area_m2: number;
  rooms: number;
  floor_level: number;
  total_floors: number;
  direction: string;
  floor_ratio: number;
  building_age_years: number;
  has_elevator: number;
  has_parking: number;
  has_pool: number;
  has_legal_paper: number;
  developer_reputation: number;
  neighborhood_score: number;
  distance_to_cbd_km: number;
  distance_to_metro_km: number;
  distance_to_school_km: number;
  distance_to_hospital_km: number;
  distance_to_park_km: number;
  distance_to_mall_km: number;
  flood_zone_risk: number;
  avg_price_district_3m_vnd_m2: number;
  listing_density: number;
  absorption_rate: number;
  dom_avg: number;
  price_momentum_30d: number;
  yoy_change: number;
  renovation_score: number;
  view_quality: number;
  interior_quality: number;
  noise_level: number;
  natural_light: number;
  month: number;
  district: string;
  price_vnd: number;
}

type TrainingRow = RawTrainingRow;

interface RetrainResult {
  model_version: string;
  metrics: {
    mae: number;
    mape: number;
    rmse: number;
    r2: number;
  };
  training_samples: number;
  validation_samples: number;
  test_samples: number;
  best_params: Record<string, unknown>;
}