BAML — Type-Safe AI Function Framework

What it is

BAML (Basically A Made-up Language) is a framework that adds engineering rigor to prompt engineering. It provides type-safe structured outputs from LLMs with guaranteed schema compliance. You define the shape of your expected output as types, write your prompt in a .baml file, and BAML ensures the LLM response conforms to your schema. It supports Python, TypeScript, Ruby, and Java.

BAML targets AI application developers who need reliable structured data from LLMs, not free-form text. Whether you are extracting entities, classifying text, or generating structured records, BAML guarantees the output matches your type definition.

Why it saves time or tokens

Parsing unstructured LLM output is fragile. JSON mode helps but still produces schema violations. BAML handles parsing, validation, and retry logic automatically. When the LLM returns malformed output, BAML fixes it or retries with a corrective prompt. This eliminates the custom parsing code and error handling that every AI application otherwise needs.

How to use

1. Install BAML: pip install baml-py or npm install @boundaryml/baml
2. Define your types and functions in .baml files
3. Generate client code and call the functions from your application

Example

// classify.baml
class Sentiment {
  label string @description('positive, negative, or neutral')
  confidence float @description('0.0 to 1.0')
  reasoning string
}

function ClassifySentiment(text: string) -> Sentiment {
  client GPT4o
  prompt #"
    Analyze the sentiment of the following text.
    Text: {{ text }}
    {{ ctx.output_format }}
  "#
}

# Generated Python client
from baml_client import b

result = b.ClassifySentiment('This product is amazing')
print(result.label)       # 'positive'
print(result.confidence)  # 0.95

Related on TokRepo

• AI tools for coding — AI development frameworks on TokRepo
• Prompt library — prompt engineering tools and patterns

Common pitfalls

• BAML's type system is separate from your language's types; keep the .baml definitions in sync with your application data models
• Complex nested types with optional fields increase the chance of LLM parsing failures; keep schemas as flat as possible
• The code generation step must run after every .baml file change; integrate it into your build pipeline to avoid stale clients