Please read “Writing AI functions”, “Testing AI functions”, and “BAML Project Structure” to get started - the rest of the documentation will only make sense if you’ve read all three articles!


Follow the installation instructions and run baml init in a new project.

The starting project structure will look like this:


Before you call an LLM, ask yourself what kind of input or output youre expecting. If you want the LLM to generate text, then you probably want a string, but if you’re trying to get it to collect user details, you may want it to return a complex type like UserDetails.

Thinking this way can help you decompose large complex prompts into smaller, more measurable functions, and will also help you build more complex workflows and agents.

We’ll start with a simple function to extract verbs from a sentence, and then build on that to learn how BAML can modify more complex and powerful functions.

Implementing an AI function

1. Define AI functions and models in BAML files

First we will define a function of the following signature in BAML: ExtractVerbs(input: string) -> string[]

Here’s the BAML equivalent, which you can add to your main.baml:

function ExtractVerbs {
    input string
    /// list of verbs
    output string[]

Every BAML function has a strictly typed input and output. The input and output can be either a primitive type (string, number, boolean) or a complex type (think unions, lists, or even custom pydantic models)

2. Implement the function using a prompt

To implement the function we need two things:

  1. An LLM client that defines which LLM to call and with which params.
  2. The actual prompt.

Define the LLM client

To implement a client we can just define one like this in a BAML file. Learn more about clients and non-openai chat providers.

If you used baml init you should already have a clients.baml file with the client below

client<llm> GPT4 {
  provider baml-openai-chat
  options {
    model gpt-4 
    api_key env.OPENAI_API_KEY

Use any parameters available to that model, like temperature etc, by adding them to the options block. You can also use environment variables to store secrets like API keys.


Define a prompt

Next we can create the prompt by implementing the function using an LLM. In BAML we provide helper utilities to inject the input variables into the prompt, and also get the LLM to return the right output type. You always get full-view of the whole prompt string, without any magic.

impl<llm, ExtractVerbs> version1 {
  client GPT4
  prompt #"
    Extract the verbs from this INPUT:
    {// this is a comment inside a prompt! //}
    Return a {#print_type(output)}.


In VSCode you can click on “Open Playground” on top of the impl or prompt to see the full prompt:

In here you’ll notice how our language automatically dedents strings, injects variables into the prompt, and supports comments that will be stripped from the actual prompt. See our syntax guide for more information on basic string / comment syntax.

We will explain more how print_type works in later tutorials.

3. Use the function in your application

Our VSCode extension automatically generates a baml_client in your language of choice - either Python or TypeScript - to access and call your functions.

from baml_client import baml as b
import asyncio

async def main():
  verb_list = await b.ExtractVerbs("This is a paragraph")

  if len(verb_list) == 1:
      print("There is 1 verb in this paragraph")
      print(f"There are {len(verb_list)} verbs in this paragraph")

if __name__ == "__main__":

Show me the code

Here it is! Clone the repo to get syntax highlighting.

Further reading