BAML

BAML is a domain-specific programming language designed to build reliable AI workflows and agents. It transforms prompts into structured functions with defined inputs and outputs, making prompt engineering more modular, testable, and maintainable.

Every prompt in BAML is a function that takes parameters and returns a specific type. BAML uses a schema-aligned parsing (SAP) algorithm to ensure outputs conform to your defined schema, even if the underlying LLM (Large Language Model) returns messy or unstructured text. You can also chain BAML functions to create complex agents and workflows.

The key features of BAML include type-safe outputs, which ensures structured and validated responses from any LLM (like JSON, XML, or YAML). It is modular and composable, allowing you to build agents as small, testable functions. BAML is language agnostic, working with any LLM (such as OpenAI, Google Gemini, or Deepseek) and any programming language. It offers testing and debugging with built-in support for unit tests, CI/CD pipelines, and runtime assertions. Additionally, BAML provides automatic retry and fallback to handle API errors gracefully, and streaming support to get type-safe outputs even when streaming partial results.

BAML differs from regular prompt engineering by focusing on output schemas rather than just crafting prompt text. It provides programmatic control over LLM outputs, reducing the need for re-prompting and manual parsing. Furthermore, BAML functions are testable and maintainable, similar to regular software code.

Yes. BAML supports any LLM provider (OpenAI, Google Gemini, Deepseek, etc.) and can be configured with custom API keys and retry policies.

To get started with BAML, you should install the BAML SDK and VSCode extension, define your functions and schemas in BAML syntax, use the BAML CLI or playground to test and run your agents, and finally integrate BAML outputs into your application code.

Yes, BAML is open source and available on GitHub.

You can find documentation and examples through the official documentation, the GitHub repository, and video tutorials.

BAML is designed for AI engineers, MLOps engineers, data scientists, and developers who are building AI agents, chatbots, or structured workflows.

Yes. BAML supports streaming structured data with type-safe outputs for each chunk.

BAML handles errors and retries by automatically retrying failed LLM API calls, and it allows you to define fallback responses and custom retry policies.

Yes. BAML allows dynamic schema updates, which is useful for self-optimizing or autonomous agents.