metadspy

Metadspy: The framework for specifying―not programming―language models.

Generate complete, reusable DSPy programs by writing simple YAML specifications. It's that easy!

---

Installation

pip install metadspy

Quickstart

1. Write a spec (spec.yaml):

   signature:
     name: SentimentClassifierSignature
     docstring: |
       Classify the overall sentiment (positive, negative or neutral) of a short text.
     inputs:
       - name: text
         kind: text
         desc: "User-provided text to analyse"
     outputs:
       - name: sentiment
         kind: choices
         choices: [positive, negative, neutral]
         desc: "Predicted sentiment label"
       - name: confidence
         kind: float
         desc: "Model confidence in [0, 1]"
     instructions: >
       Respond with one sentiment label and a confidence score between 0 and 1.
   
   module:
     name: SentimentClassifier
     type: Predict
     use: SentimentClassifierSignature
   
   llm:
     name: openai/gpt-4
     provider: openai
     api_key_env: OPENAI_API_KEY
     temperature: 0.0
     max_tokens: 1000

2. Generate your DSPy program from the CLI:

   metadspy spec.yaml --out sentiment_module.py

3. Inspect the generated code (sentiment_module.py):

   import dspy
    from dotenv import load_dotenv
    from typing import Literal
   
    load_dotenv()
   
    class SentimentClassifierSignature(dspy.Signature):
        """Classify the overall sentiment (positive, negative or neutral) of a short text."""
        text: str = dspy.InputField(desc="User-provided text to analyse")
        sentiment: Literal['positive', 'negative', 'neutral'] = dspy.OutputField(desc="Predicted sentiment label")
        confidence: float = dspy.OutputField(desc="Model confidence in [0, 1]")
   
    SentimentClassifierSignature = SentimentClassifierSignature.with_instructions(
        "Respond with one sentiment label and a confidence score between 0 and 1.\n"
    )
   
    llm = dspy.LM(
        model="openai/gpt-4o-mini",
        model_type="chat",
        temperature=0.0,
        max_tokens=1000,
        **{}
    )
   
    dspy.configure(lm=llm)
   
    SentimentClassifier = dspy.Predict(
        SentimentClassifierSignature
    )

4. Run it: You will need to add something like:

   print(SentimentClassifier(text="DSPy is so nice."))

   Then:

   python sentiment_program.py

   You’ll get a Prediction(sentiment=…, confidence=…) object back automatically.

---

Python API

If you prefer to script it:

from metadspy.parser import load_spec
from metadspy.generator import generate_code

# Load and validate
spec = load_spec("spec.yaml")

# Render and write the DSPy program
generate_code(spec, output_path="my_dspy_program.py")

---

Why Metadspy?

• Declarative: Focus on what you want, not how to do it.
• Modular: Swap LLMs, modules (Predict, ReAct, etc.), or optimizers by editing YAML.
• Reproducible: YAML + code generation = consistent, shareable pipelines.
• Extensible: Drop in custom modules, callbacks, assertions, or optimization strategies.

---

Give it a try. Specify once, ship DSPy code everywhere! 🚀 We are sooooo baaaaack !

MIT License

Project still in early stages. Enjoy !