OpenAI Setup Guide

​

Overview

​

Getting an OpenAI API Key

​

Step 1: Create an OpenAI Account

1. Visit OpenAI Platform
2. Sign up for an account or log in
3. Navigate to the API Keys page

​

Step 2: Generate an API Key

1. Click “Create new secret key”
2. Name your key (e.g., “AI-BIM App”)
3. Copy the key immediately - you won’t be able to see it again
4. Store it securely

​

Step 3: Configure Your Application

VITE_OPENAI_API_KEY=sk-proj-xxxxxxxxxxxxxxxxxxxx

​

Setting Up Billing

​

Prepaid Credits

1. Go to Billing Settings
2. Click “Add payment method”
3. Add a credit card or set up auto-recharge
4. Set a usage limit to control costs

​

Setting Usage Limits

1. Navigate to Usage Limits
2. Set a monthly budget cap
3. Configure email alerts at specific thresholds (e.g., 50%, 75%, 100%)

​

Understanding Token Usage and Costs

​

What Are Tokens?

• 1 token ≈ 4 characters in English
• 1 token ≈ ¾ of a word
• Both input (prompt) and output (response) count toward usage

​

Current Model Pricing

​

Cost Estimation for BIM Queries

1. System prompt (~50 tokens)
2. IFC file content (varies, potentially thousands of tokens)
3. User question (~10-50 tokens)

• Input: ~10,100 tokens = $0.005
• Output: ~100 tokens = $0.00015
• Total per query: ~$0.005

​

Model Selection

​

GPT-3.5 Turbo (Default)

model: "gpt-3.5-turbo"

• Fast response times
• Cost-effective for high-volume queries
• Good for straightforward BIM data extraction
• Context window: 16,385 tokens

• Material queries
• Element counting
• Simple property lookups
• Budget-conscious deployments

​

GPT-4 and GPT-4 Turbo

model: "gpt-4-turbo"  // or "gpt-4"

• More accurate understanding
• Better reasoning for complex relationships
• Higher cost (20-60x more expensive)
• Context window: 128,000 tokens (gpt-4-turbo)

• Complex spatial relationships
• Multi-step reasoning
• Large building models
• Quality-critical applications

​

API Implementation Details

​

Current API Call Structure

const response = await fetch("https://api.openai.com/v1/chat/completions", {
  method: "POST",
  headers: {
    "Content-Type": "application/json",
    "Authorization": `Bearer ${apiKey}`
  },
  body: JSON.stringify({
    model: "gpt-3.5-turbo",
    messages: [
      { 
        role: "system",  
        content: `Based on the given data answer question about ifc file.
                You should only create the response based on the information given.
                Information that is not found on given file should not be presented.
                Please answer the questions using as few words as possible`
      },
      { 
        role: "user", 
        content: `Here is the file content:\n${this.fileData}\n\nNow answer this question: ${message} using only given content` 
      }
    ]
  })
});

​

Data Optimization

modifyDataDile = () => {
  const validENtities = new Set([
    "IFCMATERIAL",
    "IfcRelAssociatesMaterial",
    "IfcBuildingStorey",
    "IfcRelContainedInSpatialStructure",
    "IFCWALL",
    "IFCSLAB",
    "IFCBEAM",
  ]);
  // Filters IFC file to only include relevant entities
}

​

Best Practices for Prompt Engineering with BIM Data

​

System Prompt Optimization

1. Only use provided data
2. Answer concisely
3. Not fabricate information

role: "system",
content: `You are a BIM data assistant. Analyze the provided IFC file data and answer questions accurately.

Rules:
- Only reference data explicitly present in the file
- Use technical BIM terminology correctly
- Respond concisely with specific values
- If information is not found, state "Not found in file data"
- When counting elements, provide exact numbers
- Include relevant IFC entity types in responses`

​

Query Structure Tips

• “How many IFCWALL elements are in the building?”
• “List all materials associated with slabs”
• “What building storeys are defined?”
• “Count beams on level 2”

• Open-ended questions
• Requests for design recommendations
• Queries about data not in IFC files
• Very complex multi-part questions

​

Token Optimization Strategies

1. Filter file content before sending (use modifyDataDile() method)
2. Send only relevant entities for the query type
3. Implement response caching for repeated queries
4. Use shorter system prompts
5. Encourage concise user questions

​

Context Window Management

• Reserve ~50 tokens for system prompt
• Reserve ~50 tokens for user question
• Reserve ~500 tokens for response
• Available for IFC data: ~15,785 tokens (~63,000 characters)

​

Monitoring and Debugging

​

Usage Dashboard

1. Visit OpenAI Usage Dashboard
2. View token consumption by date
3. Analyze cost trends
4. Download usage reports

​

Response Logging

console.log("gpt response: ", responseData);

​

Error Handling

​

Advanced Configuration

​

Adding Parameters

body: JSON.stringify({
  model: "gpt-3.5-turbo",
  messages: [...],
  temperature: 0.3,        // Lower = more deterministic
  max_tokens: 500,         // Limit response length
  top_p: 1,                // Nucleus sampling
  frequency_penalty: 0,    // Reduce repetition
  presence_penalty: 0      // Encourage diversity
})

​

Recommended Settings for BIM Queries

{
  temperature: 0.2,      // High accuracy, low creativity
  max_tokens: 300,       // Concise responses
  top_p: 0.9            // Focus on high-probability tokens
}

​

Security Considerations

​

Production Architecture Recommendation

Client (Browser)
    ↓
Backend API (Your Server)
    ↓
OpenAI API

• API key exposure
• Unauthorized usage
• Billing abuse
• CORS issues

​

Related Documentation

• Environment Variables
• ChatGpt Component Reference
• OpenAI API Documentation