onux 0.1.0

Keras-style symbolic DAGs for LLM programs

onux

Keras-style symbolic DAGs for LLM programs.

onux is an early Python library prototype inspired by the ideas in keras-dspy.md: semantic inputs, composable layers, and model graphs that look and feel like Keras, but target LLM workflows.

Install

uv add onux

For local development:

uv venv .venv
source .venv/bin/activate
uv sync

Quick example

from onux import Input, Model
from onux.layers import ChainOfThought, Generate

question = Input("question")
context = Input("context", type=list[str])

answer = ChainOfThought("answer")([question, context])
score = Generate(("score", float))([question, answer])

model = Model(
    inputs=[question, context],
    outputs=score,
    name="qa_pipeline",
)

model.compile(optimizer="auto_prompt", meta_lm="gpt-4o")
model.fit(
    [
        {
            "question": "What's the capital of France?",
            "context": ["Paris is the capital of France."],
            "score": 1.0,
        }
    ]
)

model.summary()

Current status

This package is intentionally minimal right now. It includes:

• symbolic Input(...)
• a base Layer
• graph-closing Model(...)
• placeholder built-in layers like Generate, ChainOfThought, ReAct, Retrieve, ExecuteSQL, and Map

The current release is mainly for package reservation and initial API scaffolding.

License

MIT