Skip to main content

​
Overview

Functions are BAML-powered LLM functions that you can:
  • Create and edit with structured inputs and outputs
  • Run evaluations against test cases
  • Re-run with different models and configurations
  • Version and publish for production use

​
Creating a Function

  1. Navigate to Functions in the main navigation
  2. Click Create Function
  3. Enter a name for your function (e.g., ExtractName)
  4. Define your function using BAML syntax

​
Example Function

class NameOutput {
  firstName string
  lastName string
}

function ExtractName(text: string) -> NameOutput {
  client "openai/gpt-4o"
  prompt #"
    Extract the first and last name from the following text:

    {{ text }}

    Return the name as JSON.
  "#
}

​
Function Editor

The function editor provides:
  • Syntax Highlighting: BAML syntax is highlighted for readability
  • Error Checking: Syntax errors are shown inline
  • Auto-save: Changes are saved automatically as drafts
  • Version History: View and restore previous versions

​
Function Tabs

Each function has multiple views:
TabDescription
EditorWrite and edit your BAML code
BehaviorsRun evaluations and manage test cases
VersionsBrowse and restore previous versions

​
Behaviors (Evaluations)

The Behaviors tab is where you run evaluations against your function:

​
Test Cases

Add test cases to evaluate your function:
  1. Go to the Behaviors tab
  2. Click Add Test Cases
  3. Add inputs as JSON or import from captures
[
  { "text": "My name is John Doe" },
  { "text": "Jane Smith is a software engineer" },
  { "text": "Dr. Robert Johnson, PhD" }
]

​
Running Evaluations

  1. Select test cases to run
  2. Choose which function version to evaluate
  3. Click Run
  4. Review outputs and compare with expected results

​
Importing from Captures

Build test cases from real production data:
  1. Go to Labeling and find relevant captures
  2. Label them with appropriate tags
  3. Import labeled captures as test cases

​
Versioning

Every time you save a function, a new version is created:
  • Draft: Unsaved changes (not deployed)
  • Published: The active version used by the SDK
  • Previous Versions: Historical versions you can restore

​
Publishing Changes

  1. Make changes in the editor
  2. Click Save to create a draft
  3. Click Publish to make the version active

​
Restoring a Version

  1. Open the Versions tab
  2. Find the version you want to restore
  3. Click Restore
  4. Review and publish when ready

​
BAML Syntax

Functions are defined using BAML. Here are the key concepts:

​
Classes

Define structured output types: 
class Person {
  firstName string
  lastName string
  age int?  // optional field
  email string?
}

​
Enums

Define categorical outputs: 
enum Sentiment {
  POSITIVE
  NEGATIVE
  NEUTRAL
}

​
Functions

Define the function signature and prompt: 
function ClassifySentiment(text: string) -> Sentiment {
  client "openai/gpt-4o"
  prompt #"
    Classify the sentiment of the following text as POSITIVE, NEGATIVE, or NEUTRAL:

    {{ text }}
  "#
}

​
Clients

Configure which LLM provider to use: 
client "openai/gpt-4o"        // OpenAI GPT-4o
client "anthropic/claude-3"    // Anthropic Claude 3
client "openai/gpt-3.5-turbo" // OpenAI GPT-3.5 Turbo

​
Best Practices

  • Descriptive Names: Use clear, descriptive function names like ExtractInvoiceData instead of Extract
  • Typed Outputs: Always define a class for your output to ensure type safety
  • Clear Prompts: Write clear, specific prompts with examples when needed
  • Build Evaluations: Create test cases from real captures to ensure quality
  • Version Control: Test before publishing new versions