simpledspy 0.3.1

A simplified DSPy interface

SimpleDSPy

SimpleDSPy is a lightweight Python library that simplifies building and running DSPy pipelines with an intuitive interface.

Note: This project is currently a work in progress and is not officially affiliated with or endorsed by the DSPy project. It is an independent effort to create a simplified interface for working with DSPy pipelines.

Important: Always import this package as import simpledspy, not as import dspy. The latter will import the actual DSPy library instead of SimpleDSPy.

Features

• Predict and ChainOfThought Modules: Easy-to-use function calls for predictions
• Pipeline Manager: Create complex pipelines of DSPy modules
• Optimization: Use DSPy teleprompters to optimize your modules
• Evaluation: Evaluate outputs on a scale of 1-10 and log inputs/outputs
• Reward Tracking: Track cumulative discounted rewards over time
• Advice Generation: Generate optimization advice from positive/negative examples
• CLI: Command-line interface for running modules and pipelines
• Automatic Module Creation: Generate DSPy modules from input/output specifications
• Type Hinting: Support for flexible type annotations
• Logging: Built-in logging of inputs/outputs for training data collection

Installation

pip install simpledspy

Configuration

Before using SimpleDSPy, configure your language model:

import dspy
dspy.configure(lm=dspy.OpenAI(model='gpt-3.5-turbo'))

Usage

Basic Prediction

from simpledspy import predict

# Basic prediction with default parameters
result = predict("Hello, world!")
print(result)

# Set LLM parameters directly in the call
result = predict("Write a short greeting", 
                lm_params={"max_tokens": 50, "temperature": 0.7})
print(result)

Chain of Thought

from simpledspy import chain_of_thought

# Basic chain of thought with default parameters
result = chain_of_thought("What is the capital of France?", 
                         description="Reason step by step")
print(result)

# Set temperature directly in the call
result = chain_of_thought("Solve this complex math problem", 
                         description="Show reasoning steps",
                         lm_params={"temperature": 0.3})
print(result)

Setting LLM Parameters

You can set LLM parameters directly in module calls using the lm_params argument. Supported parameters include:

• max_tokens: Maximum number of tokens to generate
• temperature: Sampling temperature (0.0-1.0)
• top_p: Nucleus sampling probability
• n: Number of completions to generate
• stop: Stop sequences
• ...and other model-specific parameters

Example:

result = predict("Generate creative text", 
                lm_params={
                    "max_tokens": 100,
                    "temperature": 0.9,
                    "top_p": 0.95
                })

Pipeline Management

from simpledspy import PipelineManager, predict

manager = PipelineManager()
manager.register_step(
    inputs=["text"],
    outputs=["cleaned"],
    module=predict("Clean text")
)
manager.register_step(
    inputs=["cleaned"],
    outputs=["sentiment"],
    module=predict("Analyze sentiment")
)

pipeline = manager.assemble_pipeline()
result = pipeline(text="I love this product!")
print(result.sentiment)  # "positive"

Optimization

from simpledspy import predict
from simpledspy.optimization_manager import OptimizationManager

# Create training data
trainset = [
    {"input": "2+2", "output": "4"},
    {"input": "3*3", "output": "9"}
]

# Optimize the module
manager = OptimizationManager()
optimized_predict = manager.optimize(predict, trainset)

# Use optimized module
result = optimized_predict("4*4")
print(result)  # "16"

Reward Tracking and Advice Generation

from simpledspy.evaluator import Evaluator

# Create evaluator with reward group
evaluator = Evaluator("Rate output quality 1-10", reward_group="math")
score = evaluator.evaluate(
    {"question": "2+2"},
    {"answer": "4"}
)
print(score)  # 10

# Get cumulative reward
cumulative = evaluator.get_cumulative_reward()
print(f"Cumulative reward: {cumulative:.2f}")

# Generate advice from examples
advice = evaluator.get_advice()
print(f"Optimization advice: {advice}")

Evaluation

from simpledspy.evaluator import Evaluator

evaluator = Evaluator("Rate output quality 1-10")
score = evaluator.evaluate(
    {"question": "2+2"},
    {"answer": "4"}
)
print(score)  # 10

CLI Usage

# Basic prediction
python -m simpledspy.cli "Hello, world!" --description "Extract greeting"

# Chain of thought reasoning
python -m simpledspy.cli "What is 2+2?" --module chain_of_thought --description "Reason step by step"

# Enable optimization
python -m simpledspy.cli "Solve math problem" --optimize --trainset data/trainset.json

# Evaluate outputs
python -m simpledspy.cli "What is the capital of France?" --evaluation-instruction "Check if answer is correct"

# Build pipelines
python -m simpledspy.cli "Input text" --pipeline "Clean text" "Analyze sentiment" "Summarize"

How It Works

SimpleDSPy provides:

1. High-level API: predict() and chain_of_thought() functions
2. Pipeline Management: Multi-step DSPy pipelines
3. Optimization: Module optimization with training data
4. Evaluation: 1-10 scoring with custom instructions
5. Logging: Input/output logging for training data
6. Reward Function: evaluate() method for cumulative rewards
7. Training Data: Extract high-scoring examples from logs

Contributing

Contributions are welcome! Please open an issue or pull request on GitHub.

Publishing the Package

To publish a new version of SimpleDSPy:

1. Update the version in pyproject.toml and simpledspy/__init__.py

2. Build the package:

   poetry build
   

3. Publish using Poetry's built-in command:

   poetry publish
   

   Poetry will prompt you for your PyPI credentials

4. Verify the published package:

   pip install --index-url https://test.pypi.org/simple/ --extra-index-url https://pypi.org/simple simpledspy==0.3.0
   

   Replace 0.3.0 with your actual version number

For API token authentication:

poetry config pypi-token.pypi your-api-token

Then run poetry publish without credentials prompt

License

MIT License