dspy

Define a metric function and training examples, then compile your module with an optimizer like BootstrapFewShot or MIPRO. The optimizer generates few-shot demonstrations and improved instructions from your data, producing a module that scores higher on your metric.

Predict makes a direct LM call from a signature without reasoning steps, making it faster for simple tasks. ChainOfThought generates a rationale before the answer, roughly doubling latency but improving accuracy on reasoning and math problems.

DSPy optimizes prompts automatically from training data and offers typed signatures with high modularity, suiting complex multi-stage systems. LangChain fits quick prototypes and simple chains using existing tool integrations where manual prompting is acceptable.

Yes, DSPy supports local models through dspy.OllamaLocal pointed at a local base URL such as http://localhost:11434. It also supports OpenAI, Anthropic Claude, and multiple models configured for different pipeline stages.

BootstrapFewShot works with 10-50 examples, while MIPRO benefits from 50-200 examples plus a separate validation set. BootstrapFinetune for weight fine-tuning recommends 100 or more examples to avoid overfitting.

Overfitting happens when max_bootstrapped_demos is set too high or the training set is too small and unrepresentative. Keep bootstrapped demos at 3-5, use diverse examples covering edge cases, and always evaluate on a held-out test set.