DSPy

Programming—not prompting—LMs

Free

Description

DSPy is a declarative framework designed for constructing modular AI software. It shifts the focus from manual prompt engineering to programming AI behavior using structured Python code. The framework enables developers to define components through natural language signatures and modules, which specify input/output behavior and invoke language models (LMs).

This approach allows for building reliable and maintainable AI systems, such as classifiers, Retrieval-Augmented Generation (RAG) pipelines, or agent loops. DSPy includes optimizers that automatically tune the prompts and weights of these modules based on defined metrics and training data. This compilation process adapts the system to specific LMs and tasks, making AI software more portable and consistently effective across different models and strategies.

Key Features

Declarative Programming: Define AI behavior using Python modules and signatures, not fragile prompt strings.
Modular Composition: Build complex AI systems by combining reusable, pre-built or custom modules (e.g., ChainOfThought, ReAct).
Automatic Optimization: Compile programs using optimizers (e.g., MIPROv2, BootstrapFinetune) to tune prompts and LM weights for specific metrics.
LM Agnostic: Supports various language models including OpenAI, Anthropic, Databricks, local LMs (via Ollama, SGLang), and LiteLLM providers.
Signatures for I/O: Clearly define input/output types and constraints for modules.
Integrated Evaluation: Provides built-in metrics (e.g., SemanticF1, answer_exact_match) and evaluation tools.