DSPy

DSPy

DSPy is the open-source prompt-programming framework from Stanford that treats LLM interactions as programmable modules rather than hand-tuned prompts. DSPy lets developers compose multi-step LLM pipelines using typed signatures, then optimize prompts and (optionally) model weights against task-specific metrics. DSPy has grown to 25K+ GitHub stars and is used by production agent teams at Cursor, Mistral, Databricks, and other AI companies for systematic prompt engineering.

Production credibility: 25,000+ GitHub stars; led by Omar Khattab (Stanford PhD, now Databricks) with contributors from Stanford NLP, Together AI, and the broader open-source community. Apache 2.0 licensed. Used in production by agent and RAG teams at Cursor, Mistral, Databricks, and multiple AI labs. Multiple peer-reviewed papers including the foundational DSPy paper at ICLR 2024.

Key Features

• Typed signatures define LLM input/output contracts as Python code
• Modules compose multi-step LLM pipelines (ChainOfThought, ReAct, ProgramOfThought, custom)
• Optimizers automatically tune prompts against metrics — BootstrapFewShot, MIPRO, COPRO, others
• Optional weight optimization through LoRA-style fine-tuning of underlying models
• Model-agnostic backend supporting OpenAI, Anthropic, Mistral, local Llama, and most providers
• Apache 2.0 open-source; 25K+ GitHub stars; led by Omar Khattab (ex-Stanford, now Databricks)
• Published in peer-reviewed venues including ICLR 2024 (foundational DSPy paper)

Ideal Use Case

Production agent and RAG teams that have outgrown hand-tuned prompts — particularly engineering teams maintaining complex LLM pipelines where prompt drift, model switches, or task variation make manual prompt engineering unscalable.

How DSPy differentiates

LangChain and LlamaIndex treat prompts as strings; the developer iterates manually. DSPy treats prompts as parameters of programmable modules that get optimized automatically against task metrics. The trade-off is a steeper learning curve — DSPy's compositional model takes longer to learn than string-template LangChain. But for production pipelines where prompt brittleness and model-switching pain are the binding constraints, DSPy's optimizer-driven approach removes the manual tuning work that dominates LLM engineering effort.

FAQ

Q: What is DSPy? A: DSPy is an open-source prompt-programming framework from Stanford that treats LLM interactions as programmable modules with typed signatures, then optimizes prompts and weights automatically against task metrics.

Q: Who built DSPy? A: Omar Khattab (Stanford PhD, now at Databricks) leads DSPy with contributors from Stanford NLP, Together AI, and the broader open-source community. The foundational DSPy paper was published at ICLR 2024.

Q: DSPy vs LangChain vs LlamaIndex? A: LangChain and LlamaIndex treat prompts as strings the developer iterates manually. DSPy treats prompts as parameters that get optimized automatically against metrics. Steeper learning curve; lower manual-tuning burden in production.

Q: Is DSPy open source? A: Yes — DSPy is Apache 2.0 licensed with 25,000+ GitHub stars. The project is led by Omar Khattab with broad community contribution.

Q: Who uses DSPy in production? A: Agent and RAG teams at Cursor, Mistral, Databricks, and multiple AI labs use DSPy in production for prompt and weight optimization.

tl;dr

DSPy is the Stanford prompt-programming framework that automates prompt and weight optimization. 25K+ GitHub stars; Apache 2.0. Led by Omar Khattab. Used in production by Cursor, Mistral, Databricks for systematic agent and RAG pipeline engineering. Treats prompts as optimized parameters, not hand-tuned strings.

Related

Looking for more options? Browse the Developer Tools directory or read our best AI coding tools listicle. DSPy is also tracked on Crunchbase.