dspy

Define a metric function and training examples, then use an optimizer like BootstrapFewShot to compile your module. The optimizer generates few-shot demonstrations from your data and embeds them into prompts, improving accuracy without manual prompt editing.

Predict makes a direct LM call for simple tasks, while ChainOfThought generates reasoning steps before the answer. ChainOfThought is roughly twice as slow but significantly more accurate on math and logical reasoning tasks.

DSPy optimizes prompts automatically from training data and offers typed signatures with high modularity, while LangChain relies on manual prompt engineering with prebuilt chains. Choose DSPy when you have training data and need systematic improvement.

Yes, DSPy supports local models through dspy.OllamaLocal by pointing to a local server URL. It also works with OpenAI, Anthropic Claude, and allows mixing multiple models for different pipeline stages.

BootstrapFewShot works with 10-50 examples, while MIPRO performs best with 50-200 examples plus a separate validation set. BootstrapFinetune for model fine-tuning recommends 100 or more examples.

Avoid DSPy for quick one-off prototypes where manual prompting is faster, or when you lack any training data for optimization. It also has a steeper learning curve than simple prompt templates for basic tasks.