dspy-bootstrap-fewshot

Few-shot demonstrations can be auto-generated using a teacher model to bootstrap optimal examples from your labeled data. This Skill automatically selects and generates high-quality demonstrations from 10-50 labeled examples, compiling them into your DSPy program without manual selection.

Bootstrap few-shot optimization is ideal when manual demonstration selection becomes tedious, you have 10-50 labeled examples available, and you need demonstrations with reasoning traces for fast prompt optimization. It scales better than hand-picking demos and produces consistent results.

The optimizer requires a DSPy program (dspy.Module), a training set of labeled examples (list[dspy.Example]), and a metric function to evaluate quality. Optional parameters include max_bootstrapped_demos and max_labeled_demos to control the output size.

Yes, you can set max_bootstrapped_demos and max_labeled_demos parameters to limit the number of generated and labeled demonstrations. This lets you balance prompt quality against model latency and token cost in production.

Bootstrapped demonstrations are generated by a teacher model from your training examples, while labeled demonstrations are hand-annotated examples from your dataset. The optimizer combines both types to create the most effective few-shot prompts.

Yes, a teacher model setting controls the quality and reasoning depth of generated demonstrations. Configuring the teacher LM before compiling ensures the optimizer produces demonstrations with appropriate complexity for your task.