n8n blueprint para imobiliária: WhatsApp + CRM + listings IA com Qdrant e Ollama

  Lead enters via WhatsApp (CTA from ads, web, FB ads)
        │
        ▼
  ┌────────────────────────────────────────────────┐
  │ n8n flow 1: WhatsApp customer service          │
  │   Webhook WA ─► extract intent ─► routing      │
  │        ├─ "looking for a house" ─► flow-search │
  │        ├─ "how much is X"  ─► flow-price       │
  │        ├─ "I want to visit" ─► flow-schedule   │
  │        └─ other ─► flow-handoff                │
  └────────────────────────────────────────────────┘
        │
        ▼
  ┌────────────────────────────────────────────────┐
  │ flow-search                                    │
  │   ├─ parse query ─► slots {city, range, beds, type}
  │   ├─ Qdrant.search hybrid over properties      │
  │   ├─ rerank top-20 → 5                         │
  │   └─ send WA cards with image + link           │
  └────────────────────────────────────────────────┘
        │
        ▼
  ┌────────────────────────────────────────────────┐
  │ flow-schedule                                  │
  │   ├─ MCP calendar.find_slot                    │
  │   ├─ confirm slot with user                    │
  │   ├─ create_event + add to CRM                 │
  │   └─ confirm via WhatsApp + Email agent        │
  └────────────────────────────────────────────────┘
        │
        ▼
  ┌────────────────────────────────────────────────┐
  │ n8n flow 2: Auto indexing                      │
  │   Cron 30 min ─► query CRM / Google Sheet      │
  │     ├─ new properties ─► Ollama describes      │
  │     ├─ embed and upsert to Qdrant              │
  │     └─ publish to channels (web, portal, FB)   │
  └────────────────────────────────────────────────┘
        │
        ▼
  ┌────────────────────────────────────────────────┐
  │ n8n flow 3: Weekly lead recalibration          │
  │   Cron Sunday 22:00 ─► read CRM contacts       │
  │     ├─ LLM scoring ─► hot/warm/cold            │
  │     ├─ to hot → seq 3 WA messages              │
  │     ├─ to warm → email seq                     │
  │     └─ metrics to Langfuse                     │
  └────────────────────────────────────────────────┘

services:
  n8n:
    image: n8nio/n8n:1.70
    restart: unless-stopped
    environment:
      DB_TYPE: postgresdb
      DB_POSTGRESDB_HOST: langfuse-db
      DB_POSTGRESDB_DATABASE: n8n
      DB_POSTGRESDB_USER: lf
      DB_POSTGRESDB_PASSWORD: ${LF_DB_PASSWORD}
      N8N_ENCRYPTION_KEY: ${N8N_ENCRYPTION_KEY}
      N8N_HOST: n8n.${DOMAIN}
      WEBHOOK_URL: https://n8n.${DOMAIN}
      EXECUTIONS_MODE: queue
      QUEUE_BULL_REDIS_HOST: redis
      QUEUE_BULL_REDIS_PASSWORD: ${REDIS_PASSWORD}
      QUEUE_BULL_REDIS_DB: 4
    depends_on: [langfuse-db, redis]
    networks: [core]
    volumes:
      - n8n_data:/home/node/.n8n

{
  "nodes": [
    {
      "type": "n8n-nodes-base.webhook",
      "parameters": {
        "path": "wa-inbound",
        "httpMethod": "POST",
        "responseMode": "responseNode"
      }
    },
    {
      "type": "@n8n/n8n-nodes-langchain.agent",
      "name": "Intent Router",
      "parameters": {
        "systemPrompt": "You are an intent classifier for a real estate agency. Classify the message as: search, price, schedule, contact_agent, other. Return JSON {intent: string, slots: object}",
        "model": "litellm/claude-haiku",
        "temperature": 0
      }
    },
    {
      "type": "n8n-nodes-base.switch",
      "parameters": {
        "rules": [
          {"value": "search", "output": 0},
          {"value": "price", "output": 1},
          {"value": "schedule", "output": 2},
          {"value": "contact_agent", "output": 3}
        ]
      }
    }
  ]
}

const { city, bedrooms, max_price, property_type } = $json.slots;

const filter = {
  must: [
    { key: "city", match: { value: city } },
    { key: "bedrooms", range: { gte: bedrooms || 1 } },
    { key: "price", range: { lte: max_price || 100000000 } },
  ]
};

const embeddedQuery = await $('Embedder').first().json.embedding;

const body = {
  vector: embeddedQuery,
  filter,
  limit: 20,
  with_payload: true,
};

return body;

// node: "Rerank"
const items = $input.all();
const query = $('Webhook').first().json.text;

const pairs = items.map(i => ({ query, text: i.payload.description }));
const resp = await $helpers.httpRequest({
  method: "POST",
  url: "http://ollama:11434/api/generate",
  body: {
    model: "bge-reranker:v2-m3",
    stream: false,
    input: pairs,
  },
  json: true,
});

return resp.scores
  .map((s, i) => ({ ...items[i], score: s }))
  .sort((a, b) => b.score - a.score)
  .slice(0, 5);

// node: "Generate Description"
const prop = $json;

const prompt = `You are a real estate copywriter writing in Mexican Spanish.
Generate an 80-120 word description, warm and professional tone, for this listing:
Type: ${prop.type}
Neighborhood: ${prop.neighborhood}
City: ${prop.city}
Bedrooms: ${prop.bedrooms}
Bathrooms: ${prop.bathrooms}
m²: ${prop.m2}
Price: ${prop.price}
Amenities: ${prop.amenities.join(", ")}

Rules:
- Don't exaggerate. Don't use empty superlatives.
- Mention 2 concrete differentiators.
- End with a soft CTA.
- No emojis.`;

const resp = await $helpers.httpRequest({
  method: "POST",
  url: "http://ollama:11434/api/generate",
  body: { model: "llama3.3:8b-instruct-q4_K_M", prompt, stream: false },
  json: true,
});
return { ...prop, description: resp.response.trim() };

// node: "Embed & Upsert to Qdrant"
const emb = await embed(prop.description + " " + prop.title);
await qdrant.upsert("properties", [{
  id: prop.id,
  vector: emb,
  payload: {
    title: prop.title,
    description: prop.description,
    city: prop.city,
    bedrooms: prop.bedrooms,
    bathrooms: prop.bathrooms,
    price: prop.price,
    url: prop.url,
    images: prop.images,
    indexed_at: new Date().toISOString(),
  }
}]);

Given the contact profile and interaction history:
{contact_json}

Last exchange:
{last_messages}

Classify as hot / warm / cold and return JSON {score, reason, next_action}.