Otaro

Otaro builds on previous work for programming language models, such as DSPy and Ell.

Highlights:

• Configs built for sharing and readability: Define tasks and agents via YAML config or code
• Automatic learning: Configs are optimized under the hood without explicit examples
• Built-in deployment: Quickly deploy a task via otaro config.yml with OpenAPI support
• Smart parsing: Parse LLM outputs in a noise-resistant manner with error correction

Usage

uv add otaro

Example

Create a YAML config file:

poet.yml

# A sample config for a haiku writing task

model: gemini/gemini-2.0-flash-001

inputs:
- topic

outputs:
- haiku

And a Python script that loads the config.

poet.py

from otaro import Task

task = Task.from_config("poet.yml")
response = task.run(topic="frog")
print(response.haiku)

Then either run the python script:

uv run poet.py
Green skin on blue pond,
A croaking song fills the air,
Summer's gentle kiss.

Or host the config as an API with interactive OpenAPI docs available at http://127.0.0.1:8000 for testing.

uv run otaro poet.yml
INFO: Started server process [2387024]
INFO: Waiting for application startup.
INFO: Application startup complete.
INFO: Uvicorn running on http://127.0.0.1:8000 (Press CTRL+C to quit)

License

This project is licensed under the MIT license.