DSPy

DSPy is a framework that optimizes LLM prompts and pipelines programmatically instead of through manual prompt engineering. You define what you want (typed input/output signatures), and DSPy's optimizers automatically find the best prompts, few-shot examples, and configurations.

Key use cases include:

• Automatic prompt optimization — replace hand-tuned prompts with data-driven optimization
• Building modular LLM pipelines with typed, composable modules
• Few-shot example selection — automatically choosing the best examples for in-context learning
• Systematic model comparison — evaluate the same pipeline across different LLMs
• Fine-tuning data generation — compile optimized prompts into fine-tuning datasets
• Academic NLP research with reproducible LLM experiments

DSPy is used by ML engineers and researchers who want to move beyond trial-and-error prompt tweaking to systematic, reproducible LLM program optimization. It is developed at Stanford NLP and has gained strong adoption in research and production settings.

The key insight: treat LLM prompts as learnable parameters that can be optimized on a small dataset, just like neural network weights.

1. Install DSPy:

   pip install dspy

2. Configure your LLM:

   import dspy
   lm = dspy.LM('openai/gpt-4o', api_key='sk-...')
   dspy.configure(lm=lm)

3. Define a module and optimize:

   class QA(dspy.Signature):
       question: str = dspy.InputField()
       answer: str = dspy.OutputField()
   
   qa = dspy.ChainOfThought(QA)
   
   optimizer = dspy.BootstrapFewShotWithRandomSearch(metric=my_metric)
   optimized_qa = optimizer.compile(qa, trainset=examples)

4. Read the DSPy documentation for advanced optimizers and multi-module pipelines.

Pricing: DSPy is free and open source (MIT). Optimization requires LLM API calls (cost depends on dataset size and number of optimization trials). Works with all major LLM providers.