Faithful port of DSPy 3.2.1 KNNFewShot (dspy/teleprompt/knn_fewshot.py).

Construction embeds the trainset once through the required :vectorizer (via Imp.Predict.KNN, the DSPy KNN port). The compiled program then, on EVERY forward call, retrieves the k nearest trainset examples for that call's inputs and runs a full Imp.Optimizer.BootstrapFewShot compilation of the student over those neighbors — metric/teacher-driven bootstrapped demonstrations, not raw attached neighbors — before executing the compiled student on the inputs (upstream KNNFewShot.compile's patched forward_pass).

few_shot_bootstrap_args maps DSPy's **few_shot_bootstrap_args onto Imp.Optimizer.BootstrapFewShot.new/2: pass :metric plus any BootstrapFewShot options (:max_bootstrapped_demos, :max_labeled_demos, :max_rounds, ...). A teacher passes to compile/3 as teacher:, mirroring upstream's compile(student, teacher=...).

Successful predictions are annotated with the retrieved neighbors under prediction.metadata.knn_few_shot (an Imp observability extension; the prompt-visible behavior is upstream's).

Example

knn_few_shot =
  Imp.Optimizer.KNNFewShot.new(3, trainset,
    vectorizer: Imp.Embeddings.BagOfWords,
    few_shot_bootstrap_args: [metric: metric]
  )

program = Imp.Optimizer.KNNFewShot.compile(knn_few_shot, student)