Dynamic Few-Shot Selection — Retrieve Relevant Examples

I will implement dynamic few-shot example selection — building an example library with semantic embeddings, and selecting the most relevant examples for each query at inference time based on input similarity, rather than using the same static examples for every query. Static few-shot examples are a compromise — the same examples cannot be equally relevant for every input in a diverse distribution; dynamic selection that retrieves the examples most similar to the current input consistently outperforms static selection on tasks with high input variability, because the model is shown the most relevant demonstrations for each specific input it needs to process.

The implementation covers example library construction with embeddings, similarity-based retrieval at inference time, example count optimisation for context window budget, diversity-balanced retrieval preventing redundant examples, and performance comparison against static few-shot.