/create-function

Create BAML Function

You are tasked with creating a new BAML function definition based on user requirements.

Task

Create a BAML function as specified in "$ARGUMENTS". Parse the requirements and scaffold a complete function definition.

Steps

1. Parse requirements from arguments:

   • Function name
   • Input parameters and types
   • Output type/structure
   • LLM provider preference
   • Prompt description

2. Determine file location:

   • Find baml_src/ directory
   • Choose appropriate .baml file (main.baml or create new one)
   • Read existing content to avoid conflicts

3. Design function structure:

   • Define input types (create custom types if needed)
   • Define output type (class, enum, or primitive)
   • Select appropriate LLM client
   • Craft initial prompt using Jinja syntax

4. Generate BAML code:

   • Write function definition
   • Include output type class if custom
   • Add inline documentation
   • Include example test case

5. Add to BAML file:

   • Append to appropriate .baml file
   • Ensure proper formatting
   • Show the created function to user

6. Next steps:

   • Suggest running /baml-toolkit:generate to create client code
   • Offer to create a test case
   • Provide usage example

Example Usage

/baml-toolkit:create-function "ExtractReceipt that takes an image and returns vendor, items, and total"
/baml-toolkit:create-function "ClassifyEmail with text input returning category (spam/important/newsletter)"
/baml-toolkit:create-function "SummarizeArticle: input is article_text (string), output is Summary class with title, key_points (array), and sentiment"

BAML Function Template

// Output type definition
class Receipt {
  vendor string
  date string @description("Format: YYYY-MM-DD")
  items ReceiptItem[]
  total float
}

class ReceiptItem {
  name string
  quantity int
  price float
}

// Function definition
function ExtractReceipt(image: image) -> Receipt {
  client GPT4Vision
  prompt #"
    Extract receipt information from the provided image.

    Return a structured receipt with:
    - Vendor/merchant name
    - Purchase date (format YYYY-MM-DD)
    - List of items with quantities and prices
    - Total amount

    {{ _.role('user') }}
    Image: {{ image }}
  "#
}

// Test case
test ExtractReceipt {
  functions [ExtractReceipt]
  args {
    image @"./test_images/sample_receipt.jpg"
  }
  assert {
    {
      checks [
        this.vendor.length > 0,
        this.total > 0,
        this.items.length > 0
      ]
    }
  }
}

Common Patterns

Classification:

enum EmailCategory {
  SPAM
  IMPORTANT
  NEWSLETTER
  PERSONAL
}

function ClassifyEmail(text: string) -> EmailCategory {
  client GPT4
  prompt #"Classify this email: {{ text }}"#
}

Extraction:

class ContactInfo {
  name string
  email string?
  phone string?
}

function ExtractContact(text: string) -> ContactInfo {
  client Claude
  prompt #"Extract contact information from: {{ text }}"#
}

Generation:

function GenerateSummary(article: string) -> string {
  client GPT4
  prompt #"Summarize in 2-3 sentences: {{ article }}"#
}

Notes

• Use ? suffix for optional fields (e.g., email string?)
• Use @description for field documentation
• Jinja syntax for dynamic prompts: {{ variable }}
• Available clients: GPT4, GPT4Vision, Claude, ClaudeVision, Gemini, etc.
• Output can be: primitives (string, int, float, bool), classes, enums, arrays, maps