An embedding-based nearest-neighbor retriever over a trainset, with the same query and ranking semantics as DSPy 3.2.1 KNN.

Construction embeds every trainset example ONCE through the required :vectorizer (any Imp.Embeddings provider — a module implementing embed/2 or a 2-arity function; DSPy's Embedder analog). Each example is rendered as "key: value" pairs for its INPUT fields joined by " | ", exactly like upstream's trainset_casted_to_vectorize. At call time the query inputs are rendered the same way, embedded, and scored against the trainset by dot product; the top k examples return in descending-score order (upstream's argsort()[-k:][::-1]).

Two details matter when moving data between Python and Elixir:

• Field iteration order: Python dicts iterate in insertion order; Elixir maps do not preserve insertion order. Pass keyword-list inputs (or single-input-field examples) when the exact multi-field rendering order matters for embedding equality.
• Tie-breaking among equal scores follows a stable sort (higher original index wins within a tie, matching a stable argsort); NumPy's default quicksort leaves ties unspecified.

For token-overlap retrieval without an embedding provider, use Imp.Retrieve.Memory.