Repo or source
https://github.com/stanfordnlp/dspy
What to explain
Instruction optimization as search: MIPROv2 (dspy/teleprompt/mipro_optimizer_v2.py) runs in three acts — bootstrap few-shot demo candidates, have a prompt_model propose instruction candidates grounded in the program, data, and traces (propose/), then use Optuna Bayesian search over instruction×demo combinations, scoring minibatches with the metric and promoting winners to full evaluations. Cover the auto='light'/'medium'/'heavy' budget presets and the separate prompt_model vs task_model roles. The insight: prompt engineering becomes hyperparameter search over a space the optimizer writes for itself.
This book is part of the series "DSPY" (book 7 of 8) — pass --series "DSPY" --series-order 7 to video.mjs.
Cover title
MIPROv2
Subtitle (optional)
Propose, Search, Select
Cover animal (optional)
honeybee
Accent color (optional hex)
#14b8a6
Model
(claude-fable-5)
Repo or source
https://github.com/stanfordnlp/dspy
What to explain
Instruction optimization as search: MIPROv2 (dspy/teleprompt/mipro_optimizer_v2.py) runs in three acts — bootstrap few-shot demo candidates, have a prompt_model propose instruction candidates grounded in the program, data, and traces (propose/), then use Optuna Bayesian search over instruction×demo combinations, scoring minibatches with the metric and promoting winners to full evaluations. Cover the auto='light'/'medium'/'heavy' budget presets and the separate prompt_model vs task_model roles. The insight: prompt engineering becomes hyperparameter search over a space the optimizer writes for itself.
This book is part of the series "DSPY" (book 7 of 8) — pass --series "DSPY" --series-order 7 to video.mjs.
Cover title
MIPROv2
Subtitle (optional)
Propose, Search, Select
Cover animal (optional)
honeybee
Accent color (optional hex)
#14b8a6
Model
(claude-fable-5)