dspy-miprov2-optimizer

Bayesian optimization jointly tunes both instructions and few-shot demonstrations for DSPy programs, exploring the instruction-demo space systematically to find higher-performing combinations. MIPROv2 uses a three-stage Bootstrap-Propose-Search workflow to ground candidates in training data and converge on state-of-the-art results.

Bayesian optimization is essential for large datasets (200+ training examples) and extended optimization runs (40+ trials) where both instructions and demos require tuning. Manual approaches become inefficient at this scale; Bayesian methods achieve better performance with fewer manual iterations.

Yes. MIPROv2 supports RAG workflows and any DSPy program with sufficient training data. Configure your language model, define your agent, and pass it to the optimizer with your trainset and metric. It outputs a fully compiled, optimized DSPy module ready for deployment.

The optimizer requires a training dataset (200+ examples recommended), a DSPy program, and a metric function to evaluate performance. Larger datasets and longer trial counts (40+) enable better exploration of the instruction-demo space and higher-quality results.

Bootstrap generates diverse candidate instructions and demonstrations from training data. Propose grounds them in examples. Search explores the combined space efficiently using Bayesian optimization, converging faster than random search or greedy approaches while maintaining exploration breadth.

Tuning only instructions limits expressiveness; jointly optimizing instructions and demonstrations allows the system to adapt both the guidance and the examples shown to the model. This combined approach yields state-of-the-art performance impossible to achieve by tuning either component in isolation.