DSPy: Declarative Language Model Programming

Stanford NLP's framework for programming—not prompting—language models.

Quick Start

import dspy

# 1. Configure
dspy.settings.configure(lm=dspy.OpenAI(model='gpt-4o-mini'))

# 2. Define Module
qa = dspy.ChainOfThought("question -> answer")

# 3. Run
response = qa(question="What is the capital of France?")
print(response.answer)

Learning Path (DAG)

The DSPy framework follows a natural progression from core concepts through optimization to advanced applications. Use this directed acyclic graph to understand dependencies and navigate the skill components.

Foundation Layer (Start Here)

1. •

   Configuring Language Models

   • •Prerequisites: None
   • •Next: Signatures, Modules, Datasets
2. •

   Designing Signatures

   • •Prerequisites: LM Configuration
   • •Next: Modules, Optimization
3. •

   Building Modules

   • •Prerequisites: Signatures
   • •Next: Optimization, Applications
4. •

   Creating Datasets

   • •Prerequisites: None
   • •Next: Optimization

Optimization Layer

1. •

   Few-Shot Learning

   • •Prerequisites: Modules, Datasets
   • •Techniques: LabeledFewShot, BootstrapFewShot, KNNFewShot
   • •Next: Applications
2. •

   Instruction Optimization

   • •Prerequisites: Modules, Datasets
   • •Techniques: COPRO, MIPROv2, GEPA
   • •Next: Applications
3. •

   Finetuning Models

   • •Prerequisites: Modules, Datasets
   • •Techniques: BootstrapFinetune
   • •Next: Applications
4. •

   Ensemble Strategies

   • •Prerequisites: Multiple trained modules
   • •Next: Applications

Application Layer

1. •

   Building RAG Pipelines

   • •Prerequisites: Modules, Optimization (recommended)
2. •

   Evaluating Programs

   • •Prerequisites: Modules, Datasets
3. •

   Integrating Haystack

   • •Prerequisites: Modules, Haystack knowledge

Advanced Features (Cross-Cutting)

1. •

   Assertions & Validation

   • •Prerequisites: Modules
2. •

   Typed Outputs

   • •Prerequisites: Signatures
3. •

   Multi-Chain Comparison

   • •Prerequisites: ChainOfThought module

Reference Documentation

• •Modules Reference - Complete module catalog
• •Optimizers Reference - All optimization techniques
• •Examples Reference - Real-world implementations

Common Workflows

Workflow 1: Basic QA System

1. •Configure LM → Design Signature → Build Module
2. •Path: configuring-language-models.md → designing-signatures.md → building-modules.md

Workflow 2: Optimized RAG System

1. •Configure LM → Build RAG Module → Optimize with Few-Shot → Evaluate
2. •Path: configuring-language-models.md → building-rag-pipelines.md → few-shot-learning.md → evaluating-programs.md

Workflow 3: Production Agent

1. •Configure LM → Design Signature → Build ReAct Module → Add Assertions → Optimize Instructions → Evaluate
2. •Path: configuring-language-models.md → designing-signatures.md → building-modules.md → assertions-validation.md → instruction-optimization.md → evaluating-programs.md

Installation

pip install dspy
# Or with specific providers
pip install dspy[anthropic]  # Claude
pip install dspy[openai]     # GPT
pip install dspy[all]        # All providers

Additional Resources

• •Official Docs: dspy.ai
• •GitHub: github.com/stanfordnlp/dspy