Optimizers Reference¶

Complete guide to DSPy optimizers used in DSPydantic.

Auto-Selection¶

DSPydantic automatically selects an optimizer based on example count:

Examples	Optimizer	Characteristics
1-2	MIPROv2 zero-shot	No examples needed, one-shot optimization
3-19	BootstrapFewShot	Few-shot learning, builds demonstrations
20+	BootstrapFewShotWithRandomSearch	More search, better with many examples

Optimizer Comparison¶

Optimizer	Speed	Quality	API Calls	Best For
MIPROv2	Fast	Fair	Low	Few examples
BootstrapFewShot	Medium	Good	Medium	Standard use
BootstrapFewShotWithRandomSearch	Slow	Excellent	High	Many examples, quality focus
COPRO	Slow	Excellent	High	Complex tasks
GEPA	Medium	Good	Medium	Balanced
SIMBA	Fast	Fair	Low	Fast iteration
BetterTogether	Slow	Excellent	High	Ensemble
BootstrapFinetune	Very Slow	Excellent	Very High	Fine-tuning

When to Use Each¶

Development/Quick Iteration: - MIPROv2 (fastest) - SIMBA (fast) - BootstrapFewShot (balanced)

Production/Quality Focus: - BootstrapFewShotWithRandomSearch (recommended) - COPRO (excellent results) - BetterTogether (ensemble)

Fine-tuning: - BootstrapFinetune (creates training data)

Configuration¶

Default configuration (auto-selected):

result = prompter.optimize(examples=examples)
# Automatically picks best optimizer for example count

Manual selection:

result = prompter.optimize(
    examples=examples,
    optimizer="BootstrapFewShot"
)

Fine-grained control with optimizer_kwargs and compile_kwargs:

result = prompter.optimize(
    examples=examples,
    optimizer="miprov2zeroshot",
    optimizer_kwargs={"auto": None, "num_candidates": 3},
    compile_kwargs={"num_trials": 5, "minibatch": False},
)

API Call Estimates¶

For 10 examples with a 5-field model:

Optimizer	Baseline	Optimization	Total
MIPROv2	~2	~15	~17
BootstrapFewShot	~2	~40	~42
BootstrapFewShotWithRandomSearch	~2	~80	~82
COPRO	~2	~120	~122

Actual costs depend on model, field count, and optimization duration.

Performance Tips¶

Reduce API calls: - Use fewer examples - Use single-pass mode (default, sequential=False) - Use a cheaper model during optimization - Use include_fields to focus on critical fields - Use early_stopping_patience in sequential mode - Limit trials with compile_kwargs={"num_trials": 5}

Improve quality: - Add more diverse examples - Use a stronger model - Use sequential=True for field-by-field optimization - Use auto_generate_prompts=True to add system/instruction prompts - Choose quality-focused optimizer