Baml Ai

Domain-specific language and toolchain for defining type-safe LLM function calls. BAML files declare input/output types, prompt templates, and LLM client configuration. baml-cli codegen emits Python, TypeScript, and Ruby clients with fully typed function signatures and built-in structured output parsing. Automatic retry and fallback logic on parse failures. Boundary Studio for prompt observability and trace inspection. Confidence is docs-derived.

BAML DSL: function ExtractOrder(input: string) -> OrderSchema { client gpt4o; prompt #'...' }; class OrderSchema { items Item[]; total float }; baml-cli generate emits b.ExtractOrder(input) in Python/TypeScript; async and sync variants; streaming support for partial structured output; test {} blocks in BAML for prompt unit testing; Boundary Studio client for trace export.

LLM provider API keys injected via environment variables referenced in BAML client config (e.g., env.OPENAI_API_KEY) — never hardcoded in .baml files; generated client uses these at runtime; Boundary Studio tracing uses a separate BAML_API_KEY for telemetry; no user data sent to Boundary ML unless Studio tracing is explicitly configured.

Structured output parsing handles both JSON-mode and raw text LLM responses via BAML's custom parser — no dependency on provider-specific structured output features; automatic retry with same or fallback LLM on parse failure; confidence score per parsed field available in extended response; TypeScript client throws typed ParseError on unrecoverable parse failures; streaming emits partial type-checked objects.

Documentation covers BAML language quickstart, type system reference (primitives, classes, enums, unions), prompt template syntax, multi-provider client config, codegen setup (Python, TypeScript, Ruby, OpenAPI), test blocks, and Boundary Studio integration. VS Code extension with BAML syntax highlighting and inline test runner. Confidence is docs-derived.