BAML Integration Skill

Universal patterns for working with BAML (Boundary ML) in any project. BAML provides type-safe LLM prompting with automatic code generation for Python and TypeScript.

Design Principle

This skill is framework-generic. It provides universal BAML patterns that work in any codebase:

• NOT tailored to CodeGraph-DE, Book-Vetting, or any specific project
• Covers common patterns applicable across all BAML projects
• Specific domain types should go in project-specific skills

Variables

Variable	Default	Description
BAML_SRC	baml_src	Directory containing BAML files
AUTO_GENERATE	true	Auto-run baml-cli generate on changes
STRICT_TYPES	true	Enforce strict type matching

Instructions

MANDATORY - Follow the Workflow steps below in order.

1. Understand BAML's role in the project
2. Check existing BAML schema and types
3. Follow type-safe patterns when working with LLMs
4. Keep generated code in sync

Red Flags - STOP and Reconsider

If you're about to:

• Define LLM prompts without BAML types
• Manually parse LLM output instead of using BAML
• Skip running baml-cli generate after schema changes
• Ignore type errors in generated clients

STOP -> Define BAML types -> Generate client -> Then proceed

Workflow

1. Understand Project BAML Setup

Check the BAML configuration:

# Find BAML source directory
find . -name "*.baml" -type f | head -5

# Check BAML client
ls -la baml_client/ || ls -la baml_src/baml_client/

# Check for generator config
cat baml_src/generators.baml 2>/dev/null

2. Review Existing Types

Before adding new types, review what exists:

// Common patterns in baml_src/types/

// Enums
enum TaskStatus {
  PENDING
  IN_PROGRESS
  COMPLETED
  FAILED
}

// Classes (DTOs)
class UserRequest {
  query string
  context string?
  preferences map<string, string>?
}

class UserResponse {
  answer string
  confidence float
  sources string[]
}

3. Define New Types

When adding LLM-powered features:

// 1. Define input type
class MyInput {
  field1 string @description("Clear description")
  field2 int @description("What this number represents")
}

// 2. Define output type
class MyOutput {
  result string
  metadata MyMetadata?
}

class MyMetadata {
  confidence float
  reasoning string
}

// 3. Define the function
function MyFunction(input: MyInput) -> MyOutput {
  client GPT4
  prompt #"
    Given: {{ input.field1 }}
    Count: {{ input.field2 }}

    Provide your analysis.

    {{ ctx.output_format }}
  "#
}

4. Generate Client

After schema changes:

# Generate Python and TypeScript clients
baml-cli generate

# Or with specific config
baml-cli generate --config baml_src/generators.baml

5. Use Generated Client

# Python usage
from baml_client import b

async def process_request(input_data: dict):
    result = await b.MyFunction(
        input=MyInput(
            field1=input_data["query"],
            field2=input_data["count"]
        )
    )
    return result.result

// TypeScript usage
import { b } from './baml_client';

async function processRequest(inputData: Record<string, unknown>) {
  const result = await b.MyFunction({
    field1: inputData.query as string,
    field2: inputData.count as number
  });
  return result.result;
}

Cookbook

Schema Synchronization

• IF: Adding or modifying BAML types
• THEN: Read and execute ./cookbook/schema-sync.md

DTO Generation

• IF: Creating data transfer objects
• THEN: Read and execute ./cookbook/dto-generation.md

Client Wrapper Patterns

• IF: Wrapping BAML client for your service
• THEN: Read and execute ./cookbook/client-wrapper.md

Quick Reference

BAML Type Syntax

Type	Syntax	Example
String	string	name string
Int	int	count int
Float	float	score float
Boolean	bool	active bool
Optional	type?	nickname string?
Array	type[]	tags string[]
Map	map<K, V>	metadata map<string, string>
Enum	enum Name	status TaskStatus
Class	class Name	Custom types
Union	type1 \| type2	result string \| Error

Function Attributes

Attribute	Purpose	Example
@description	Field documentation	@description("User's email")
@alias	JSON key mapping	@alias("user_id")
@skip	Exclude from output	@skip

Client Selection

// Define clients in clients.baml
client GPT4 {
  provider openai
  options {
    model "gpt-4-turbo"
    temperature 0.7
  }
}

client Claude {
  provider anthropic
  options {
    model "claude-3-opus"
    max_tokens 4096
  }
}

// Use in functions
function MyFunc(input: Input) -> Output {
  client GPT4  // or Claude
  prompt #"..."#
}

Retry and Fallback

// Configure retries
client GPT4WithRetry {
  provider openai
  retry_policy {
    max_retries 3
    strategy exponential_backoff
  }
}

// Fallback chain
client_fallback MainClient {
  primary GPT4
  fallback [Claude, GPT35Turbo]
}

Best Practices

1. Type Safety First

Always define explicit types:

// Good: Explicit types
class BookAnalysis {
  title string
  author string
  summary string @description("2-3 sentence summary")
  rating float @description("Rating from 0-5")
  tags string[]
}

// Bad: Using generic types
function AnalyzeBook(text: string) -> string  // Loses type safety

2. Use Descriptions

Add descriptions for LLM guidance:

class SearchQuery {
  query string @description("The user's search query in natural language")
  filters SearchFilters? @description("Optional filters to narrow results")
  limit int @description("Maximum number of results to return, default 10")
}

3. Handle Errors

Define error types:

class Error {
  code string
  message string
}

function SafeAnalysis(input: Input) -> Output | Error {
  // LLM can return either success or error
}

4. Version Your Schema

Keep schema versions aligned:

// baml_src/version.baml
// Schema version: 1.2.0
// Last updated: 2025-12-24

// Document breaking changes in CHANGELOG

Integration Points

With Schema Alignment

BAML types should align with database models:

// BAML type
class User {
  id int
  email string
  name string?
}

// Should match SQLAlchemy model
class User(Base):
    id: Mapped[int]
    email: Mapped[str]
    name: Mapped[str | None]

With API Schemas

BAML types can generate API response types:

// BAML response type
class APIResponse {
  success bool
  data ResponseData
  error string?
}

// Use generated types in FastAPI
@app.post("/analyze")
async def analyze(request: Request) -> APIResponse:
    result = await b.Analyze(request.data)
    return APIResponse(success=True, data=result)

With Frontend Types

Generated TypeScript types work with frontend:

// Generated by BAML
import type { BookAnalysis } from './baml_client/types';

// Use in React component
function BookCard({ analysis }: { analysis: BookAnalysis }) {
  return (
    <div>
      <h2>{analysis.title}</h2>
      <p>{analysis.summary}</p>
      <Rating value={analysis.rating} />
    </div>
  );
}

Troubleshooting

Generation Errors

# Check BAML syntax
baml-cli check

# Verbose generation
baml-cli generate --verbose

Type Mismatches

If LLM output doesn't match expected type:

1. Check prompt for clarity
2. Add more explicit @description hints
3. Consider using union types with Error
4. Enable strict mode in client

Client Import Issues

# Ensure client is generated
try:
    from baml_client import b
except ImportError:
    # Run: baml-cli generate
    raise RuntimeError("BAML client not generated. Run: baml-cli generate")