oddspy

orchstration device for dspy

I wanted a simple-ish workflow builder for LLM workflows and wanted to learn so wrote this - it's composable action blocks that all operate on a shared dictionary as the storage for output and memory. This has it's disadvantages but as long as it fits in memory, this seems useful to me for straightforward assembly of structured data through chains of LLM calls.

How it works

Building Blocks

You can create layers of Pipelines made of Steps.

Each Step (BaseStep or LMStep) uses Processors which define either arbitrary code execution and/or language-model use. dspy.Signatures are used in each Processor input into LMSteps.

Configuration allows for setting per-task modules and dspy.Modules such as Predict, ChainOfThought, and ReAct (soon, with tools).

Information Flow

Here are two example steps:

[
    BaseStep(
        step_type="test base step",
        processor_class=ExampleBaseProcessor,
        output_key="base_step_answer"
    ),
    LMStep(
        step_type="test lm step",
        lm_name=LMForTask.DEFAULT,
        processor_class=ExampleLMProcessor,
        output_key="lm_step_answer",
        depends_on=["*"]
    ),   
]

The output_key of the BaseStep writes to a shared dictionary which the LMStep can access, and which accumulates the outputs of all prior steps. Passing depends_on=["*"] is passing all output keys generated before this step into the current step.

Alternatively for larger workflows you could pass in depends_on=["previously_written_key", "another_key.nested_key.another_nested_key"] for precise control of which information is passed into each step (hence, the keys available for context for the LLM calls within that step).

Installation

(you should use uv) uv pip install oddspy

or

git clone *url* && uv pip install -e .

Usage

1. Define a lm_config.yaml in your root based on lm_config.yaml.example
2. Define OPENROUTER_API_KEY in your .env if using openrouter
3. Create! See examples/example_pipeline.py for a basic workflow created using this framework

Desired Features to Add Next

1. dspy.ReAct support + any other new modules
2. Checkpoint support - tracking step execution and restarting from any step
3. Improved error transparency and aesthetics for logging
4. Interactivity? Tool template if needed? Infra for code-writing/editing agents?