Why Use BAML?

​

The Problem with Current Approaches

# The old way - error-prone and hard to maintain
def home():
    return "<button onclick=\"() => alert(\\\"hello!\\\")\">Click</button>"

• F-strings and string concatenation: Hard to read, easy to break
• No type safety: Runtime errors from schema mismatches
• Slow iteration: Must execute code to test prompt changes
• Manual JSON schemas: Tedious to maintain and sync with code

1. Set up a complete Python environment
2. Execute your entire application
3. Wait for the LLM response
4. Fix the prompt
5. Repeat

​

The BAML Solution

​

1. Test 10x Faster

• See the full rendered prompt with multimodal assets
• View the exact API request being sent
• Run tests in parallel for even faster iteration
• Test in 5 seconds instead of 2 minutes

​

2. Fully Type-Safe

from baml_client import b
from baml_client.types import Resume, Education

# Autocomplete works!
resume: Resume = b.ExtractResume(text)
school: str = resume.education[0].school
year: int = resume.education[0].year

stream = b.stream.ExtractResume(text)
for partial in stream:  # partial: PartialResume
    if partial.name:  # Type-safe optional fields
        print(f"Name: {partial.name}")

​

3. Works with Any Model

function ExtractResume(text: string) -> Resume {
- client "openai/gpt-4o"
+ client "anthropic/claude-3-5-sonnet"
  prompt #"..."
}

• OpenAI (GPT-4, GPT-4o, O1, O3)
• Anthropic (Claude 3, Claude 3.5)
• Google (Gemini, Vertex AI)
• AWS (Bedrock)
• Azure OpenAI
• Any OpenAI-compatible API (Ollama, OpenRouter, VLLM, LMStudio, TogetherAI, Deepseek, etc.)

​

4. Reliable Structured Outputs

• Handles markdown within JSON
• Supports chain-of-thought before answers
• Works on Day 1 of new model releases
• No need to check if a model supports parallel tool calls, recursive schemas, anyOf, oneOf, etc.

​

5. Built-in Reliability

client<llm> MyClient {
  provider openai
  options {
    model gpt-4o
  }
  
  // Automatic retries
  retry_policy {
    max_retries 3
    strategy exponential_backoff
  }
}

// Fallback to another model
function Extract(text: string) -> Data {
  client MyClient | ClaudeBackup
  prompt #"..."
}

// Round-robin across models
function Extract(text: string) -> Data {
  client RoundRobin<MyClient, Claude, Gemini>
  prompt #"..."
}

​

6. Maintainable Code

# Hard to read, easy to break
prompt = f"""
Extract the following data from this resume:
{{
  \"name\": \"string\",
  \"education\": [{{
    \"school\": \"string\",
    \"degree\": \"string\",
    \"year\": \"number\"
  }}]
}}

Resume:
{resume_text}

Respond with valid JSON only.
"""

result = json.loads(client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": prompt}]
).choices[0].message.content)

​

BAML vs Other Frameworks

​

vs Pydantic AI / Instructor

​

vs LangChain / LlamaIndex

​

vs Raw API Calls

​

Why a New Language?

def home():
    return "<button onclick=\"() => alert(\\\"hello!\\\")\">Click</button>"

function Home() {
  return <button onClick={() => setCount(prev => prev + 1)}>
           {count} clicks!
         </button>
}

​

Design Philosophy

1. Avoid invention when possible
   • Prompts need versioning → use Git
   • Prompts need storage → use filesystems
2. Any file editor and terminal should work
   • No special tools required
   • Works with your existing workflow
3. Be fast
   • Built in Rust for maximum performance
   • So fast, you can’t even tell it’s there
4. Easy to understand
   • A first-year university student should be able to read BAML
   • Simple syntax, powerful results

​

100% Open Source & Private

• Apache 2.0 License - Use it freely in commercial projects
• No telemetry - Zero network requests beyond your explicit model calls
• Not used for training - Your prompts and data stay private
• Local-first - Works completely offline (except for model calls)
• Git-friendly - BAML files diff beautifully in version control

​

Production Ready

• Weekly updates and improvements
• Stable API with semantic versioning
• Active community on Discord
• Comprehensive documentation
• Example projects and templates

​

Common Questions

​

Do I need to write my whole app in BAML?

​

Is BAML production-ready?

​

Can I use BAML with my existing framework?

​

What if I need to switch models at runtime?

​

Get Started