[feat] Enable Dynamic Instantiation of BAML Enums + Bug Report

[dlmgary]

Product

BAML

Problem Statement / Use Case

Hi folks!

First, I want to say that I'm a huge fan of BAML and think you're on the right track to find the correct level of abstraction for AI/code invocation.

Feature Request

The feature I'm requesting (which there might already be a way to do, but I haven't been able to find) is the ability to instantiate dynamically generated BAML enums on the caller side.

Problem

I haven't found a way to dynamically instantiate the BAML enums that I create via the TypeBuilder. The enums are available to the BAML client callee when passed via baml_options={"tb": tb}, but they're not available to the caller.

Current Behavior

Given the following enum definition:

# enums.baml 
enum MyBamlEnum {
    VariableOne
    @@dynamic
}

I can add new variables dynamically:

from baml_client import b as baml_client
from baml_client.type_builder import TypeBuilder

tb = TypeBuilder()
tb.MyBamlEnum.add_value(name="VariableTwo")
baml_client.MyFunction(..., baml_options={"tb": tb})

As expected, MyFunction sees MyBamlEnum with both variables: VariableOne and VariableTwo. However, the dynamically added values are not accessible to the function caller:

tb = TypeBuilder()
tb.MyBamlEnum.add_value(name="VariableTwo")
tb.MyBamlEnum.list_values()
# Returns: [('VariableOne', <baml_py.baml_py.EnumValueBuilder at 0x147069e90>)]
# Notice "VariableTwo" is missing

Desired Behavior

Ideally, tb.MyBamlEnum would allow users to instantiate their newly created dynamic enum with something like:

tb = TypeBuilder()
tb.MyBamlEnum.add_value(name="VariableTwo")

# Creates a new class with the added variable
updated_enum = tb.MyBamlEnum.get_class()
print(updated_enum)
# Expected: <enum 'MyBamlEnum'>

print(updated_enum("VariableOne"))
# Expected: MyBamlEnum.VariableOne

print(updated_enum("VariableTwo"))
# Expected: MyBamlEnum.VariableTwo

This would be extremely helpful for downstream applications (e.g., creating mocks, validation, etc.).

Bug Report: Inconsistent Enum Return Types

Additionally, I believe there's a bug in how enums are handled by BAML functions. The enum returns different types depending on how instances were created:

• Static enum values return objects: <MyBamlEnum.VariableOne: 'VariableOne'>
• Dynamic enum values return strings: 'VariableTwo'

Reproduction Steps

Using this sample BAML function:

function MyFunction(
    name_of_variable_to_return: string  
) -> MyBamlEnum {
    client "openai/gpt-4.1-nano"
    prompt #"
    return the enum called: {{name_of_variable_to_return }}
    {{ ctx.output_format }}
    "#
}

from baml_client import b as baml_client
from baml_client.type_builder import TypeBuilder
from baml_client.types import MyBamlEnum

tb = TypeBuilder()
tb.MyBamlEnum.add_value(name="VariableTwo")

resp_1 = baml_client.MyFunction(name_of_variable_to_return="variable_one", baml_options={"tb": tb})
print(resp_1)
# Actual: <MyBamlEnum.VariableOne: 'VariableOne'>

resp_2 = baml_client.MyFunction(name_of_variable_to_return="variable_two", baml_options={"tb": tb})
print(resp_2)
# Actual: 'VariableTwo'
# Expected: <MyBamlEnum.VariableTwo: 'VariableTwo'>

Expected Behavior

Both static and dynamic enum values should return the same type of object for consistency.

[linear]

BAML-494 [feat] Enable Dynamic Instantiation of BAML Enums + Bug Report

[hellovai]

hello @dlmgary! Funnily enough, great timing. I've actually been working on a fully reflective type system at this moment.

#2483

We should be merging this in for the upcomming release!

Not only will you be able to get the full typesystem, but also attributes like @alias and @description will be available.

(note there will be a slight breaking change for this, and we'll share migration docs + have great errors for existing code!)

[dlmgary]

Love it, thanks @hellovai!