​
Build a Multi-Model AI Assistant with Cost Tracking

This guide shows you how to build a customer support assistant that intelligently routes queries to different AI models based on complexity, using Vercel AI Gateway for model access and Helicone for cost tracking and analytics.

​
Prerequisites

​
Setup

Install the required packages: 
npm install @ai-sdk/openai ai

​
Create the AI Client

Set up a client that routes through Helicone for monitoring: 
import { createOpenAI } from '@ai-sdk/openai';
import { generateText, tool } from 'ai';
import { z } from 'zod';

const ai = createOpenAI({
  apiKey: process.env.VERCEL_API_KEY,
  baseURL: 'https://gateway.helicone.ai/v1',
  headers: {
    'Helicone-Auth': `Bearer ${process.env.HELICONE_API_KEY}`,
    'Helicone-Target-URL': 'https://ai-gateway.vercel.sh/v1/ai',
    'Helicone-Target-Provider': 'VERCEL'
  }
});

​
Classify Query Complexity

Use gpt-4o-nano with tool calling for precise classification: 
import { tool } from 'ai';
import { z } from 'zod';

const classifyTool = tool({
  description: 'Classify a customer support query by complexity',
  parameters: z.object({
    complexity: z.enum(['simple', 'complex', 'technical']).describe(
      'simple: Basic questions about account, passwords, features. ' +
      'complex: Refunds, complaints, escalations, urgent issues. ' +
      'technical: API errors, integration issues, code problems.'
    ),
    reasoning: z.string().describe('Brief explanation for the classification')
  })
});

async function classifyQueryComplexity(query: string): Promise<'simple' | 'complex' | 'technical'> {
  const result = await generateText({
    model: ai('openai/gpt-4o-nano'),
    tools: {
      classify: classifyTool
    },
    toolChoice: 'required',
    prompt: `Classify this customer query: "${query}"`
  });

  // Get the classification from the tool call
  const toolCall = result.toolCalls[0];
  return toolCall.args.complexity;
}

​
Route to Appropriate Model

Use different models based on query complexity to optimize costs: 
async function handleCustomerQuery(query: string, customerId: string) {
  const complexity = await classifyQueryComplexity(query);
  
  // Track complexity in Helicone
  const headers = {
    'Helicone-User-Id': customerId,
    'Helicone-Property-Complexity': complexity,
    'Helicone-Property-Department': 'customer-support'
  };
  
  let model;
  switch (complexity) {
    case 'simple':
      model = ai('openai/gpt-4o-mini'); // Cheapest, handles basic queries
      break;
    case 'complex':
      model = ai('openai/gpt-4o'); // Better reasoning for complex issues
      break;
    case 'technical':
      model = ai('anthropic/claude-3-5-sonnet'); // Excellent for technical support
      break;
  }
  
  const response = await generateText({
    model,
    messages: [
      {
        role: 'system',
        content: 'You are a helpful customer support assistant. Be concise and professional.'
      },
      {
        role: 'user',
        content: query
      }
    ],
    headers,
    temperature: 0.3, // Lower temperature for consistent support responses
    maxTokens: 200
  });
  
  return {
    answer: response.text,
    model: complexity,
    usage: response.usage
  };
}

​
Implement Response Caching

Cache all queries regardless of complexity for maximum cost savings: 
async function handleQueryWithCache(query: string, customerId: string) {
  const complexity = await classifyQueryComplexity(query);
  
  // Enable caching for all complexity levels
  const headers = {
    'Helicone-User-Id': customerId,
    'Helicone-Property-Complexity': complexity,
    'Helicone-Cache-Enabled': 'true',
    'Helicone-Cache-Bucket-Max-Size': '10',
    'Helicone-Cache-Seed': 'support-v1'
  };
  
  // Select model based on complexity
  let model;
  switch (complexity) {
    case 'simple':
      model = ai('openai/gpt-4o-mini');
      break;
    case 'complex':
      model = ai('openai/gpt-4o');
      break;
    case 'technical':
      model = ai('anthropic/claude-3-5-sonnet');
      break;
  }
  
  return await generateText({
    model,
    messages: [
      { role: 'system', content: 'You are a helpful support agent.' },
      { role: 'user', content: query }
    ],
    headers,
    temperature: 0 // Zero temperature for consistent cache hits
  });
}

​
Complete Support System

Here’s the full implementation: 
import { createOpenAI } from '@ai-sdk/openai';
import { generateText } from 'ai';

// Initialize AI client with Helicone
const ai = createOpenAI({
  apiKey: process.env.VERCEL_API_KEY,
  baseURL: 'https://gateway.helicone.ai/v1',
  headers: {
    'Helicone-Auth': `Bearer ${process.env.HELICONE_API_KEY}`,
    'Helicone-Target-URL': 'https://ai-gateway.vercel.sh/v1/ai',
    'Helicone-Target-Provider': 'VERCEL'
  }
});

interface SupportTicket {
  id: string;
  customerId: string;
  query: string;
  priority: 'low' | 'medium' | 'high';
}

async function processSupportTicket(ticket: SupportTicket) {
  const complexity = await classifyQueryComplexity(ticket.query);
  
  // Model selection based on complexity and priority
  let model;
  if (ticket.priority === 'high' || complexity === 'technical') {
    model = ai('anthropic/claude-3-5-sonnet');
  } else if (complexity === 'complex') {
    model = ai('openai/gpt-4o');
  } else {
    model = ai('openai/gpt-4o-mini');
  }
  
  try {
    const response = await generateText({
      model,
      messages: [
        {
          role: 'system',
          content: `You are a customer support agent. Priority: ${ticket.priority}. Be helpful and professional.`
        },
        {
          role: 'user',
          content: ticket.query
        }
      ],
      headers: {
        'Helicone-User-Id': ticket.customerId,
        'Helicone-Property-TicketId': ticket.id,
        'Helicone-Property-Priority': ticket.priority,
        'Helicone-Property-Complexity': complexity,
        // Enable caching for all queries
        'Helicone-Cache-Enabled': 'true',
        'Helicone-Cache-Bucket-Max-Size': '20',
        'Helicone-Cache-Seed': 'support-v1'
      },
      temperature: 0, // Zero temperature for consistent cache hits
      maxTokens: 250
    });
    
    return {
      ticketId: ticket.id,
      response: response.text,
      model: model.modelId,
      cost: response.usage // Track in Helicone dashboard
    };
  } catch (error) {
    console.error('Support ticket processing failed:', error);
    throw error;
  }
}

// Example usage
const ticket: SupportTicket = {
  id: 'TICKET-12345',
  customerId: 'CUST-789',
  query: 'How do I reset my password?',
  priority: 'low'
};

const result = await processSupportTicket(ticket);
console.log(`Response sent to customer: ${result.response}`);

​
Monitor Performance

View your assistant’s performance in Helicone:
  1. Cost Analysis: Compare costs across different models
  2. Response Times: Monitor latency by model and complexity
  3. Cache Hit Rate: Track savings from cached responses
  4. User Analytics: See which customers need the most support
Helicone dashboard showing model usage and costs

​
Optimize Based on Data

Use Helicone’s analytics to:
  • Identify common queries for caching
  • Adjust model selection thresholds
  • Track cost per ticket complexity
  • Monitor customer satisfaction by model

​
Next Steps

Custom Properties

Track additional metadata

Caching

Reduce costs with smart caching

User Metrics

Analyze per-customer usage

Alerts

Set up cost and error alerts