docs/architecture/MARITACA_OPTIMIZATION_GUIDE.md

🚀 Guia de Otimização Maritaca AI - Cidadão.AI

Resumo das Melhorias

1. Novo Endpoint Otimizado

• URL: /api/v1/chat/optimized
• Modelo: Sabiazinho-3 (mais econômico)
• Persona: Carlos Drummond de Andrade
• Economia: ~40-50% menor custo por requisição

2. Comparação de Modelos

Modelo	Custo	Qualidade	Tempo Resposta	Uso Recomendado
Sabiazinho-3	💰	⭐⭐⭐⭐	1-5s	Conversas gerais, saudações
Sabiá-3	💰💰💰	⭐⭐⭐⭐⭐	3-15s	Análises complexas

3. Endpoints Disponíveis

# 1. Simple (Sabiá-3) - FUNCIONANDO 100%
POST /api/v1/chat/simple

# 2. Stable (Multi-fallback) - NOVO
POST /api/v1/chat/stable

# 3. Optimized (Sabiazinho-3 + Drummond) - NOVO
POST /api/v1/chat/optimized

Integração Frontend - Versão Otimizada

Serviço de Chat Atualizado

// services/chatService.ts
export interface ChatEndpoint {
  url: string;
  name: string;
  priority: number;
  model: string;
}

export class ChatService {
  private readonly API_URL = process.env.NEXT_PUBLIC_API_URL
  
  private endpoints: ChatEndpoint[] = [
    {
      url: '/api/v1/chat/optimized',
      name: 'Optimized (Sabiazinho)',
      priority: 1,
      model: 'sabiazinho-3'
    },
    {
      url: '/api/v1/chat/simple',
      name: 'Simple (Sabiá-3)',
      priority: 2,
      model: 'sabia-3'
    },
    {
      url: '/api/v1/chat/stable',
      name: 'Stable (Fallback)',
      priority: 3,
      model: 'mixed'
    }
  ]
  
  async sendMessage(
    message: string, 
    options?: {
      preferredModel?: 'economic' | 'quality';
      useDrummond?: boolean;
    }
  ): Promise<ChatResponse> {
    const sessionId = `session_${Date.now()}`
    
    // Select endpoint based on preference
    let selectedEndpoints = [...this.endpoints]
    
    if (options?.preferredModel === 'economic') {
      // Prioritize Sabiazinho
      selectedEndpoints.sort((a, b) => 
        a.model === 'sabiazinho-3' ? -1 : 1
      )
    } else if (options?.preferredModel === 'quality') {
      // Prioritize Sabiá-3
      selectedEndpoints.sort((a, b) => 
        a.model === 'sabia-3' ? -1 : 1
      )
    }
    
    // Try endpoints in order
    for (const endpoint of selectedEndpoints) {
      try {
        const body: any = { message, session_id: sessionId }
        
        // Add Drummond flag for optimized endpoint
        if (endpoint.url.includes('optimized')) {
          body.use_drummond = options?.useDrummond ?? true
        }
        
        const response = await fetch(`${this.API_URL}${endpoint.url}`, {
          method: 'POST',
          headers: { 'Content-Type': 'application/json' },
          body: JSON.stringify(body)
        })
        
        if (response.ok) {
          const data = await response.json()
          console.log(`✅ Success with ${endpoint.name}`)
          return data
        }
      } catch (error) {
        console.warn(`Failed ${endpoint.name}:`, error)
      }
    }
    
    // Ultimate fallback
    return {
      message: 'Desculpe, estou temporariamente indisponível.',
      session_id: sessionId,
      agent_name: 'Sistema',
      agent_id: 'system',
      confidence: 0,
      metadata: { fallback: true }
    }
  }
  
  // Analyze message to decide best model
  analyzeComplexity(message: string): 'simple' | 'complex' {
    const complexKeywords = [
      'analise', 'investigue', 'compare', 'tendência',
      'padrão', 'anomalia', 'detalhe', 'relatório'
    ]
    
    const hasComplexKeyword = complexKeywords.some(
      keyword => message.toLowerCase().includes(keyword)
    )
    
    return hasComplexKeyword || message.length > 100 
      ? 'complex' 
      : 'simple'
  }
}

Componente Inteligente

// components/SmartChat.tsx
export function SmartChat() {
  const [messages, setMessages] = useState<Message[]>([])
  const [modelPreference, setModelPreference] = useState<'auto' | 'economic' | 'quality'>('auto')
  const chatService = new ChatService()
  
  const handleSendMessage = async (text: string) => {
    // Add user message
    const userMessage = createUserMessage(text)
    setMessages(prev => [...prev, userMessage])
    
    // Analyze complexity for auto mode
    let preference: 'economic' | 'quality' | undefined
    
    if (modelPreference === 'auto') {
      const complexity = chatService.analyzeComplexity(text)
      preference = complexity === 'simple' ? 'economic' : 'quality'
    } else if (modelPreference !== 'auto') {
      preference = modelPreference
    }
    
    // Send with appropriate model
    const response = await chatService.sendMessage(text, {
      preferredModel: preference,
      useDrummond: true // Enable cultural persona
    })
    
    // Add response
    const assistantMessage = {
      ...createAssistantMessage(response),
      metadata: {
        ...response.metadata,
        model_preference: preference,
        actual_model: response.model_used
      }
    }
    
    setMessages(prev => [...prev, assistantMessage])
    
    // Log for monitoring
    logChatMetrics({
      model_used: response.model_used,
      response_time: response.metadata?.response_time_ms,
      tokens: response.metadata?.tokens_used,
      success: true
    })
  }
  
  return (
    <div className="smart-chat">
      {/* Model preference selector */}
      <div className="model-selector">
        <label>Modo:</label>
        <select 
          value={modelPreference} 
          onChange={(e) => setModelPreference(e.target.value as any)}
        >
          <option value="auto">Automático</option>
          <option value="economic">Econômico (Sabiazinho)</option>
          <option value="quality">Qualidade (Sabiá-3)</option>
        </select>
      </div>
      
      {/* Chat messages */}
      <MessageList messages={messages} />
      
      {/* Input */}
      <ChatInput onSend={handleSendMessage} />
      
      {/* Status indicator */}
      <ChatStatus 
        lastModel={messages[messages.length - 1]?.metadata?.actual_model}
        preference={modelPreference}
      />
    </div>
  )
}

Otimizações de Custo

1. Cache Inteligente

class CachedChatService extends ChatService {
  private cache = new Map<string, CachedResponse>()
  
  async sendMessage(message: string, options?: any) {
    // Check cache for common questions
    const cacheKey = this.normalizeMessage(message)
    const cached = this.cache.get(cacheKey)
    
    if (cached && !this.isExpired(cached)) {
      return {
        ...cached.response,
        metadata: {
          ...cached.response.metadata,
          from_cache: true
        }
      }
    }
    
    // Get fresh response
    const response = await super.sendMessage(message, options)
    
    // Cache if successful
    if (response.confidence > 0.8) {
      this.cache.set(cacheKey, {
        response,
        timestamp: Date.now()
      })
    }
    
    return response
  }
}

2. Batching de Requisições

class BatchedChatService extends ChatService {
  private queue: QueuedMessage[] = []
  private timer: NodeJS.Timeout | null = null
  
  async sendMessage(message: string, options?: any) {
    return new Promise((resolve) => {
      this.queue.push({ message, options, resolve })
      
      if (!this.timer) {
        this.timer = setTimeout(() => this.processBatch(), 100)
      }
    })
  }
  
  private async processBatch() {
    const batch = this.queue.splice(0, 5) // Max 5 per batch
    
    // Send all at once (if API supports)
    const responses = await this.sendBatch(batch)
    
    // Resolve individual promises
    batch.forEach((item, index) => {
      item.resolve(responses[index])
    })
    
    this.timer = null
  }
}

Métricas e Monitoramento

// utils/chatMetrics.ts
export class ChatMetricsCollector {
  private metrics = {
    totalRequests: 0,
    modelUsage: new Map<string, number>(),
    avgResponseTime: 0,
    totalTokens: 0,
    errorRate: 0,
    cacheHitRate: 0
  }
  
  recordMetric(data: ChatMetric) {
    this.metrics.totalRequests++
    
    // Track model usage
    const model = data.model_used || 'unknown'
    this.metrics.modelUsage.set(
      model,
      (this.metrics.modelUsage.get(model) || 0) + 1
    )
    
    // Update averages
    this.updateAverages(data)
    
    // Send to analytics (optional)
    if (window.gtag) {
      window.gtag('event', 'chat_interaction', {
        model_used: model,
        response_time: data.response_time,
        success: !data.error
      })
    }
  }
  
  getCostEstimate(): number {
    const sabiazinhoCost = 0.001 // per request
    const sabia3Cost = 0.003 // per request
    
    const sabiazinhoCount = this.metrics.modelUsage.get('sabiazinho-3') || 0
    const sabia3Count = this.metrics.modelUsage.get('sabia-3') || 0
    
    return (sabiazinhoCount * sabiazinhoCost) + (sabia3Count * sabia3Cost)
  }
  
  getReport() {
    return {
      ...this.metrics,
      estimatedCost: this.getCostEstimate(),
      modelDistribution: Object.fromEntries(this.metrics.modelUsage)
    }
  }
}

Recomendações de Uso

Para o Frontend:

1. Perguntas Simples/Saudações: Use Sabiazinho (economic mode)
2. Análises Complexas: Use Sabiá-3 (quality mode)
3. Auto Mode: Deixa o sistema decidir baseado na complexidade

Economia Estimada:

• Conversas simples: 40-50% economia usando Sabiazinho
• Mix típico (70% simples, 30% complexo): ~35% economia total
• Com cache: Adicional 10-20% economia

Próximos Passos:

1. Implementar cache para perguntas frequentes
2. Adicionar análise de sentimento para ajustar tom
3. Criar dashboards de custo em tempo real
4. A/B testing entre modelos