cdp-edge 2.1.0 → 2.2.1

package/server-edge-tracker/index.js

@@ -16,6 +16,11 @@ import {
   sha256,
   META_TO_GA4,
   VALID_EVENT_NAMES,
+  resolveFunnelStage,
+  resolveIntentScore,
+  distanceBucketWeight,
+  computeMetaSignalWeights,
+  metaSignalBucket,
 } from './modules/utils.js';
 
 // ── Banco de dados (D1) ───────────────────────────────────────────────────────
@@ -31,6 +36,7 @@ import {
   saveEdgeFingerprint,
   resurrectUTM,
   upsertLtvProfile,
+  recordLtvFeedback,
 } from './modules/db.js';
 
 // ── Dispatch — plataformas de ads ─────────────────────────────────────────────
@@ -92,6 +98,16 @@ import {
   buildGoogleCustomerMatchExport,
 } from './modules/intelligence.js';
 
+// ── Haversine distance (km) — sem dependência externa ────────────────────────
+function haversineKm(lat1, lon1, lat2, lon2) {
+  const R    = 6371;
+  const dLat = (lat2 - lat1) * Math.PI / 180;
+  const dLon = (lon2 - lon1) * Math.PI / 180;
+  const a    = Math.sin(dLat / 2) ** 2 +
+    Math.cos(lat1 * Math.PI / 180) * Math.cos(lat2 * Math.PI / 180) * Math.sin(dLon / 2) ** 2;
+  return R * 2 * Math.atan2(Math.sqrt(a), Math.sqrt(1 - a));
+}
+
 // ─────────────────────────────────────────────────────────────────────────────
 // HANDLER PRINCIPAL
 // ─────────────────────────────────────────────────────────────────────────────
@@ -175,7 +191,7 @@ export default {
       }
 
       try {
-        await env.AI.run('@cf/meta/llama-3.1-8b-instruct', {
+        await env.AI.run('@cf/ibm-granite/granite-4.0-h-micro', {
           messages: [{ role: 'user', content: 'ping' }],
           max_tokens: 1,
         });
@@ -270,6 +286,9 @@ export default {
         payload.engagementScore = behavioral_data.engagement_score ?? behavioral_data.totalScore ?? null;
         payload.intentionLevel  = behavioral_data.intention_level  ?? null;
         payload.userScore       = behavioral_data.user_score       ?? null;
+        // Sinais de engajamento profundo — usados no anti-falso-positivo e no LTV
+        payload.scrollScore     = behavioral_data.scroll_score     ?? null;
+        payload.timeLevel       = behavioral_data.time_level       ?? null;
         payload.email     = payload.email     || behavioral_data.email      || null;
         payload.phone     = payload.phone     || behavioral_data.phone      || null;
         payload.firstName = payload.firstName || behavioral_data.first_name  || behavioral_data.firstName || null;
@@ -280,6 +299,35 @@ export default {
         payload.dob       = payload.dob       || behavioral_data.dob         || null;
       }
 
+      // ── Normalização de intent_score → 0.0–1.0 ──────────────────────────
+      // Aceita string ('high'/'medium'/'low'), 0-1 ou 0-100 enviados pelo front.
+      // intent_bucket mantém a label legível para D1 e logs.
+      const intentScoreNum = resolveIntentScore(payload.intent_score);
+      if (intentScoreNum !== null) {
+        payload.intent_score  = intentScoreNum;
+        payload.intentScoreNum = intentScoreNum;
+        payload.intent_bucket = intentScoreNum >= 0.8 ? 'high'
+                              : intentScoreNum >= 0.5 ? 'medium' : 'low';
+      } else {
+        payload.intentScoreNum = null;
+      }
+
+      // ── Anti-falso-positivo ───────────────────────────────────────────────
+      // Penaliza intent se engajamento insuficiente: scroll raso E tempo curto.
+      // scroll_score < 2.0 ≈ não passou de 50% da página.
+      // time_level 'curioso' = menos de 60 segundos na página.
+      if (payload.intentScoreNum !== null) {
+        const isShallowScroll = payload.scrollScore !== null && payload.scrollScore < 2.0;
+        const isShallowTime   = payload.timeLevel === 'curioso';
+        if (isShallowScroll && isShallowTime) {
+          const penalized = Math.round(payload.intentScoreNum * 0.7 * 100) / 100;
+          payload.intentScoreNum    = penalized;
+          payload.intent_score      = penalized;
+          payload.intent_bucket     = penalized >= 0.8 ? 'high' : penalized >= 0.5 ? 'medium' : 'low';
+          payload.intent_penalized  = true;  // flag auditável — visível no D1 e logs
+        }
+      }
+
       // ── Edge Fingerprint + UTM Resurrection ──────────────────────────────
       const fingerprint = await generateEdgeFingerprint(request);
       payload.utmRestored = false;
@@ -324,6 +372,34 @@ export default {
 
       const ga4Name = META_TO_GA4[eventName] || eventName.toLowerCase();
 
+      // ── Dual-layer semantics ─────────────────────────────────────────────
+      // Meta sempre recebe o nome canônico (Schedule, Lead, etc.).
+      // Internamente o CDP usa semântica de funil precisa via FUNNEL_TAXONOMY.
+      if (payload.funnel_stage) {
+        const { depth, funnelDepth } = resolveFunnelStage(payload.funnel_stage);
+        payload.funnelDepth = funnelDepth;  // ex: 'bottom_intent', 'bottom_conversion'
+        payload.funnelLevel = depth;        // ex: 'bottom', 'conversion', 'mid'
+      }
+      if (eventName === 'Schedule' && payload.funnel_stage === 'route_click') {
+        payload.internalEvent = 'IntentToVisit';
+      }
+
+      // ── Real Estate Distance Enrichment ──────────────────────────────────
+      // Calcula distância entre o IP do usuário (via Cloudflare) e o imóvel.
+      // Ativa quando o evento carrega property_lat + property_lng (ex: Schedule de rota).
+      const propLat = parseFloat(payload.property_lat ?? payload.propertyLat);
+      const propLng = parseFloat(payload.property_lng ?? payload.propertyLng);
+      const userLat = parseFloat(request.cf?.latitude);
+      const userLng = parseFloat(request.cf?.longitude);
+      if (!isNaN(propLat) && !isNaN(propLng) && !isNaN(userLat) && !isNaN(userLng)) {
+        const distKm = haversineKm(userLat, userLng, propLat, propLng);
+        payload.distanceKm     = Math.round(distKm * 10) / 10;
+        payload.distanceBucket = distKm <  5 ? 'very_close' :
+                                 distKm < 15 ? 'close'      :
+                                 distKm < 30 ? 'nearby'     :
+                                 distKm < 60 ? 'moderate'   : 'far';
+      }
+
       // ── LTV Prediction (+ A/B Testing de Prompts) ─────────────────────────
       const LTV_EVENTS = ['Lead', 'Contact', 'Schedule', 'CompleteRegistration'];
       if (LTV_EVENTS.includes(eventName) && !payload.value) {
@@ -352,6 +428,35 @@ export default {
         }
       }
 
+      // ── LTV Feedback Loop — fecha o ciclo preditivo ──────────────────────
+      // Quando uma compra real acontece, registra o valor real e recalcula accuracy.
+      // Alimenta ltv_ab_variations.accuracy_score → autoDecideAbWinner usa isso.
+      if (eventName === 'Purchase' && payload.value > 0) {
+        ctx.waitUntil(recordLtvFeedback(env, payload.userId, payload.value));
+      }
+
+      // ── Meta Signal Score (composite, pesos dinâmicos) ───────────────────
+      // Pesos variam por profundidade de funil: fundo = comportamento pesa mais.
+      {
+        const w  = computeMetaSignalWeights(payload.funnelLevel);
+        const iW = payload.intentScoreNum ?? 0.5;
+        const lW = payload.ltvScore ? payload.ltvScore / 100 : 0.5;
+        const dW = distanceBucketWeight(payload.distanceBucket);
+        payload.metaSignal       = Math.round(((iW * w.intent) + (lW * w.ltv) + (dW * w.dist)) * 100) / 100;
+        payload.metaSignalBucket = metaSignalBucket(payload.metaSignal);  // 'hot'/'warm'/'cold'
+      }
+
+      // ── Hot Lead Trigger (timing + sinal) ────────────────────────────────
+      // Reage em tempo real: WhatsApp apenas quando comportamento + sinal justificam.
+      // Critérios: rota + muito próximo + (intent alto OU meta_signal alto)
+      //            + (janela de decisão BRT 18–22h OU sinal excepcional ≥ 0.9)
+      const hourBRT  = (new Date().getUTCHours() - 3 + 24) % 24;
+      const inWindow = hourBRT >= 18 && hourBRT <= 22;
+      const isHotLead = payload.funnel_stage === 'route_click'
+                     && payload.distanceBucket === 'very_close'
+                     && ((payload.intentScoreNum ?? 0) >= 0.8 || payload.metaSignal >= 0.85)
+                     && (inWindow || payload.metaSignal >= 0.9);
+
       // Cross-Device Graph — background
       if (env.DB && payload.userId && (payload.email || payload.phone)) {
         ctx.waitUntil(resolveDeviceGraph(env.DB, payload.userId, payload.email, payload.phone));
@@ -366,10 +471,10 @@ export default {
         sendMetaCapi(env, eventName, payload, request, ctx),
         sendGA4Mp(env, ga4Name, payload, ctx),
         sendTikTokApi(env, eventName, payload, request, ctx),
-        saveLead(env, eventName, payload, request, 'website'),
+        saveLead(env, payload.internalEvent || eventName, payload, request, 'website'),
         upsertProfile(env, eventName, payload, request),
-        ...(WHATSAPP_NOTIFY_EVENTS.includes(eventName)
-          ? [sendWhatsApp(env, eventName, payload)]
+        ...(WHATSAPP_NOTIFY_EVENTS.includes(eventName) || isHotLead
+          ? [sendWhatsApp(env, isHotLead ? 'hot_lead_intent_to_visit' : eventName, payload)]
           : []),
       ]);
 

package/server-edge-tracker/modules/db.js

@@ -443,6 +443,77 @@ export async function upsertLtvProfile(env, userId, ltv) {
   }
 }
 
+// ── recordLtvFeedback — fecha o ciclo preditivo com valor real de compra ─────
+// Chamado em background quando um Purchase chega com payload.value > 0.
+// Atualiza user_profiles + ltv_ab_assignments + ltv_ab_variations em cascata.
+export async function recordLtvFeedback(env, userId, realValue) {
+  if (!env.DB || !userId || !realValue || realValue <= 0) return;
+
+  try {
+    // 1. Busca predicted_ltv_value atual do perfil
+    const profile = await env.DB.prepare(`
+      SELECT predicted_ltv_value FROM user_profiles WHERE user_id = ?
+    `).bind(userId).first();
+
+    // accuracy = 1 - |pred-real|/real  (0–1, mesmo padrão do A/B test accuracy_score)
+    const predictedValue = profile?.predicted_ltv_value;
+    const ltv_accuracy = (predictedValue !== null && predictedValue !== undefined)
+      ? Math.max(0, Math.round((1 - Math.abs(predictedValue - realValue) / Math.max(realValue, 1)) * 100) / 100)
+      : null;
+
+    // 2. Grava valor real + accuracy no perfil
+    await env.DB.prepare(`
+      UPDATE user_profiles
+      SET real_ltv_value  = ?,
+          ltv_accuracy    = ?,
+          ltv_feedback_at = datetime('now'),
+          updated_at      = datetime('now')
+      WHERE user_id = ?
+    `).bind(realValue, ltv_accuracy, userId).run();
+
+    // 3. Fecha assignment do A/B test mais recente não convertido (janela 60 dias)
+    const assignment = await env.DB.prepare(`
+      SELECT id, variation_id, predicted_ltv
+      FROM ltv_ab_assignments
+      WHERE user_id  = ?
+        AND converted = 0
+        AND assigned_at > datetime('now', '-60 days')
+      ORDER BY assigned_at DESC
+      LIMIT 1
+    `).bind(userId).first();
+
+    if (!assignment) return;
+
+    // 3a. Marca assignment como convertido
+    await env.DB.prepare(`
+      UPDATE ltv_ab_assignments
+      SET converted    = 1,
+          real_revenue = ?,
+          converted_at = datetime('now')
+      WHERE id = ?
+    `).bind(realValue, assignment.id).run();
+
+    // 3b. Atualiza métricas acumuladas da variação (running average — safe para concorrência D1)
+    const predLtv  = assignment.predicted_ltv || 0;
+    const indivAcc = Math.max(0, 1 - Math.abs(predLtv - realValue) / Math.max(realValue, 1));
+
+    await env.DB.prepare(`
+      UPDATE ltv_ab_variations
+      SET total_purchases  = total_purchases + 1,
+          sum_real_revenue = sum_real_revenue + ?,
+          avg_real_revenue = (sum_real_revenue + ?) / (total_purchases + 1),
+          accuracy_score   = ROUND(
+            (COALESCE(accuracy_score, 0) * total_purchases + ?) / (total_purchases + 1),
+            4
+          )
+      WHERE id = ?
+    `).bind(realValue, realValue, indivAcc, assignment.variation_id).run();
+
+  } catch (err) {
+    console.error('[LTV-Feedback] recordLtvFeedback error:', err.message);
+  }
+}
+
 // ── Feedback Loop — Log de falhas e métricas de saúde ────────────────────────
 
 export async function logApiFailure(DB, platform, eventName, errorCode, errorMessage, eventId, rawPayload) {

package/server-edge-tracker/modules/dispatch/meta.js

@@ -24,6 +24,9 @@ export async function sendMetaCapi(env, eventName, payload, request, ctx) {
     eventId, pageUrl,
     value, currency,
     contentIds, contentName, contentType, numItems,
+    // Dual-layer context — funil avançado + imóveis
+    funnel_stage, distance_bucket, intent_score, intent_bucket,
+    ltvScore, ltvClass, metaSignal, metaSignalBucket: metaSignalBucketVal,
   } = payload;
 
   const phoneNorm   = normalizePhone(phone);
@@ -57,6 +60,15 @@ export async function sendMetaCapi(env, eventName, payload, request, ctx) {
     ...(contentName && { content_name: contentName }),
     ...(contentType && { content_type: contentType }),
     ...(numItems    && { num_items: parseInt(numItems) }),
+    // Contexto de funil e proximidade — enriquece matching e otimização Meta
+    ...(funnel_stage    && { funnel_stage }),
+    ...(distance_bucket && { distance_bucket }),
+    ...(intent_score    && { intent_score }),
+    ...(ltvScore    !== undefined && ltvScore    !== null && { ltv_score: ltvScore }),
+    ...(ltvClass    && { ltv_class: ltvClass }),
+    ...(metaSignal       !== undefined && metaSignal       !== null && { meta_signal: metaSignal }),
+    ...(metaSignalBucketVal && { meta_signal_bucket: metaSignalBucketVal }),
+    ...(intent_bucket    && { intent_bucket }),
   };
 
   const eventPayload = {

package/server-edge-tracker/modules/ml/fraud.js

@@ -6,13 +6,6 @@
 import { sha256, tryParseJson } from '../utils.js';
 
 // ── Listas de detecção ────────────────────────────────────────────────────────
-export const DISPOSABLE_EMAIL_DOMAINS = new Set([
-  'mailinator.com','guerrillamail.com','tempmail.com','throwaway.email',
-  'yopmail.com','sharklasers.com','guerrillamailblock.com','spam4.me',
-  '10minutemail.com','trashmail.com','maildrop.cc','fakeinbox.com',
-  'dispostable.com','getairmail.com','mailnull.com',
-]);
-
 export const DATACENTER_PATTERNS = /amazon|google|microsoft|digitalocean|linode|ovh|vultr|hetzner|contabo|cloudflare|packet|rackspace|leaseweb/i;
 
 // ── checkFraudGate — roda ANTES de qualquer processamento de evento ────────────
@@ -64,15 +57,7 @@ export async function checkFraudGate(env, request, payload) {
       result.score += 20; result.reasons.push('no_accept_language');
     }
 
-    // 6. Email descartável
-    if (email) {
-      const domain = email.split('@')[1]?.toLowerCase();
-      if (domain && DISPOSABLE_EMAIL_DOMAINS.has(domain)) {
-        result.score += 25; result.reasons.push('disposable_email');
-      }
-    }
-
-    // 7. Velocity check via KV
+    // 6. Velocity check via KV
     if (env.GEO_CACHE && ip) {
       const velKey1h = `fraud_velocity:${ip}:h`;
       const velStr   = await env.GEO_CACHE.get(velKey1h);

package/server-edge-tracker/modules/ml/ltv.js

@@ -8,6 +8,23 @@ import { extractFeatures, predictWithWeights, loadActiveWeights } from './logist
 // Cache key para o teste ativo (KV — evita hit no D1 a cada request /track)
 const AB_LTV_CACHE_KEY = 'ab_ltv_active_test';
 
+// ── Prompt especializado para imóveis ────────────────────────────────────────
+// Ativado automaticamente quando property_lat/lng estão presentes no payload.
+// Override por A/B test tem prioridade sobre este prompt.
+const REAL_ESTATE_PROMPT = `You are a real estate lead scoring expert for the Brazilian market.
+Reply ONLY with a JSON object {"adjustment": <integer between -15 and 15>} based on the lead data.
+Scoring rules (apply additively):
+- distance_km < 5:  +12 (lives nearby, buys fast)
+- distance_km 5-15: +8
+- distance_km 15-30: +3
+- distance_km > 30:  0
+- distance_km unknown: +3 (gave intent signal without geo)
+- event = Schedule or route click: +5 (physical visit intent)
+- scroll_score >= 3 AND time_level = comprador: +4 (deep engagement)
+- hour_brt between 18-22 (weekday): +3 (active decision window)
+- has_phone = true: +2 (reachable for follow-up)
+No explanation. JSON only.`;
+
 // ── predictLtv — Heurística em 5 dimensões (0-100 pts) ───────────────────────
 export async function predictLtv(env, payload, request, customSystemPrompt = null) {
   // ── Tentar modelo treinado (regressão logística real) ─────────────────────
@@ -83,6 +100,19 @@ export async function predictLtv(env, payload, request, customSystemPrompt = nul
   if (payload.phone) score += 4;
   if (payload.firstName) score += 2;
 
+  // 6. Proximidade ao imóvel físico (0–15) — apenas quando distância calculada
+  const distKm = parseFloat(payload.distanceKm ?? payload.user_distance_km ?? -1);
+  if (distKm >= 0) {
+    if      (distKm <  5)  score += 15;
+    else if (distKm < 15)  score += 10;
+    else if (distKm < 30)  score += 6;
+    else if (distKm < 60)  score += 3;
+    // > 60km: sem bônus — lead distante precisa de argumento diferente
+  } else if (payload.property_lat || payload.propertyLat) {
+    // Coords no payload mas distância não resolvida: pequeno bônus por intenção de rota
+    score += 3;
+  }
+
   score = Math.min(100, score);
 
   let ltvClass, ltvMultiplier;
@@ -102,21 +132,36 @@ export async function predictLtv(env, payload, request, customSystemPrompt = nul
   let aiAdjustment = 0;
   if (env.AI && score >= 40) {
     try {
+      const isRealEstate = !!(payload.property_lat || payload.propertyLat);
       const systemContent = customSystemPrompt ||
-        'You are a conversion rate expert. Reply ONLY with a JSON object {"adjustment": <number between -10 and 10>} based on the lead data provided. No explanation.';
+        (isRealEstate
+          ? REAL_ESTATE_PROMPT
+          : 'You are a conversion rate expert. Reply ONLY with a JSON object {"adjustment": <number between -10 and 10>} based on the lead data provided. No explanation.');
+
+      const userContext = {
+        utm_source:  payload.utmSource,
+        intention:   intentionLevel,
+        engagement:  engScore,
+        hour_utc:    hour,
+        country,
+        has_email:   !!payload.email,
+        has_phone:   !!payload.phone,
+      };
+      if (isRealEstate) {
+        userContext.event_type     = 'real_estate_schedule';
+        userContext.distance_km    = payload.distanceKm ?? payload.user_distance_km ?? 'unknown';
+        userContext.distance_bucket = payload.distanceBucket ?? 'unknown';
+        userContext.scroll_score   = payload.scrollScore || payload.scroll_score || 0;
+        userContext.time_level     = payload.timeLevel || payload.time_level || 'unknown';
+        userContext.intent_score   = payload.intent_score || 'high';
+        userContext.hour_brt       = (hour - 3 + 24) % 24; // UTC-3 aproximado
+      }
+
       const prompt = [
         { role: 'system', content: systemContent },
-        { role: 'user', content: JSON.stringify({
-          utm_source:  payload.utmSource,
-          intention:   intentionLevel,
-          engagement:  engScore,
-          hour_utc:    hour,
-          country,
-          has_email:   !!payload.email,
-          has_phone:   !!payload.phone,
-        })},
+        { role: 'user', content: JSON.stringify(userContext) },
       ];
-      const aiRes = await env.AI.run('@cf/meta/llama-3.1-8b-instruct', { messages: prompt, max_tokens: 32 });
+      const aiRes = await env.AI.run('@cf/ibm-granite/granite-4.0-h-micro', { messages: prompt, max_tokens: 32 });
       const parsed = JSON.parse(aiRes.response.trim());
       if (typeof parsed.adjustment === 'number') {
         aiAdjustment = Math.max(-10, Math.min(10, parsed.adjustment));