Baml

  • BAML is a domain-specific language to generate structured outputs from LLMs
    • with good DX
      • you can build reliable agents, chatbots, with RAG, MCP, etc.

High-level Developer Walkthru

main.baml 

class WeatherAPI {
  city string @description("the user's city")
  timeOfDay string @description("As an ISO8601 timestamp")
}

function UseTool(user_message: string) -> WeatherAPI {
  client "openai-responses/gpt-5-mini"
  prompt #"
    Extract.... {# we will explain the rest in the guides #}
  "#
}

  • Call your function in any language 

    import asyncio
from baml_client import b
from baml_client.types import WeatherAPI

def main():
    weather_info = b.UseTool("What's the weather like in San Francisco?")
    print(weather_info)
    assert isinstance(weather_info, WeatherAPI)
    print(f"City: {weather_info.city}")
    print(f"Time of Day: {weather_info.timeOfDay}")

if __name__ == '__main__':
    main()

Enable reliable tool-calling with any model

  • BAML includes schema-aligned parsing algorithm to support flexible outputs LLMs can provide
    • e.g. JSON blob, chain-of-thought, etc.

Schema Aligned Parsing

  • instead of relying on the model to strictly understand our desired format
    • write a parser that generously reads the output text
      • and applies error correction technique
        • with knowledge of the original schema
  • instead of `jsonschema`, `bamlschema` is used as a more compressed way to define schemas
    • this is because BAML is not as strict so we can omit characters

what is SAP?

  • the idea is that the model will make mistakes, and to build a parser that is robust enough to handle those mistakes
  • in the context of structured data extraction
    • we have a schema to guide us

Postel's Law coined by Jon Postel, creator of TCP/IP Be conservative in what you do, be liberal in what you accept from others

Some error correction techniques we use in SAP include:

  • Unquoted strings
  • Unescaped quotes and new lines in strings
  • Missing commas, colons, brackets, keys
  • Cast fractions to floats
  • Remove superfluous keys in objects
  • Strip yapping
  • Picking the best of many possible candidates in case LLM produces multiple outputs
  • Complete partial objects (due to streaming)