dspy

Use the BootstrapFewShot optimizer with a metric function and 10-50 training examples. Call optimizer.compile(module, trainset=trainset) and DSPy generates few-shot demonstrations from high-scoring predictions, improving your module without manual prompt edits.

Create a dspy.Module combining dspy.Retrieve(k=3) with a ChainOfThought module taking context and question as inputs. Configure a retriever like ChromadbRM, then optionally optimize the pipeline with BootstrapFewShot using question-answer training pairs.

Predict makes a direct LM call for simple tasks, while ChainOfThought generates reasoning steps before the answer, improving accuracy on math and logic tasks. ChainOfThought is roughly twice as slow but significantly better for complex reasoning.

Yes. Configure Claude with dspy.Claude(model=..., api_key=...), OpenAI with dspy.OpenAI, and local models via dspy.OllamaLocal pointing at a local server. You can also switch models per task using dspy.settings.context.

Use MIPRO when you have 50-200 training examples plus a separate validation set and want state-of-the-art instruction optimization. It runs 100+ trials over 10-30 minutes, while BootstrapFewShot is faster and better as a first attempt with smaller datasets.

Overfitting happens when max_bootstrapped_demos is set too high or training data is too small. Keep demonstrations at 3-5, use proper train/validation/test splits, and evaluate the optimized module only on a held-out test set.