DynamicFewShotPrompt: max_similar_examples=0 (top_k=0) returns all examples instead of none

[mittalpk]

Describe the bug

SimpleInMemoryExampleStore.get_examples(data, top_k=0, ...) is meant to return at most top_k examples, but top_k=0 returns every stored example instead of none, in both the recency-based path (no embedding model) and the cosine-similarity path (with an embedding model).

Root cause: Python's negative-slice quirk where -0 == 0, so list[-top_k:] with top_k=0 becomes list[-0:] == list[0:] == the full list, not an empty slice.

Two call sites hit this:

# recency-based path, no embedding model
if not self.embedding_model or not self._embeddings_list:
    return self._examples[-top_k:]   # top_k=0 -> returns ALL examples

# cosine-similarity path, inside _get_nearest_examples
if len(valid_indices) > 0:
    top_indices = valid_indices[
        np.argsort(similarities[valid_indices])[-top_k:]   # same issue
    ]

(The other fallback branch, list(range(max(0, len(embeddings) - top_k), len(embeddings))), happens to compute the empty range correctly for top_k=0 — it's specifically the two [-top_k:] slices that are wrong.)

This is reachable from the public DynamicFewShotPrompt API: DynamicFewShotPrompt(..., max_similar_examples=0) is a reasonable way to opt out of few-shot examples entirely, but instead every stored example gets silently included in the formatted prompt.

Ragas version

Reproduced against current main (298b682).

Reproduction

from ragas.prompt.dynamic_few_shot import SimpleInMemoryExampleStore

store = SimpleInMemoryExampleStore()
for i in range(5):
    store.add_example({"q": f"question {i}"}, {"a": f"answer {i}"})

result = store.get_examples({"q": "new question"}, top_k=0)
print(len(result))  # prints 5, expected 0

Expected behavior

top_k=0 returns an empty list in both code paths.

I have a fix + regression tests ready and will open a PR shortly.