[bug] Approximate search in `VectorStore` sometimes fails or returns fewer results

[MSM2002]

Description

While testing the VectorStore class, we noticed issues with the approximate search feature (enabled with use_approx=True). This feature is supposed to make searches faster by grouping similar items into clusters and searching only the closest clusters instead of the entire dataset.

However, in some cases:

1. Search fails completely with an error like:

ValueError: need at least one array to concatenate

This happens when the feature tries to look for items in a cluster but the cluster ends up empty.

2. Search returns fewer results than requested, for example asking for 2 results but getting only 1.

These problems seem related to how clusters are created and queried in the approximate search.

Steps to Reproduce

1. Create a VectorStore with a small number of items (e.g., 5 company names with 8-dimensional embeddings).
2. Build the KMeans index with 2 clusters.
3. Call search with use_approx=True and small n_probe_clusters values.

Example failing test cases:

def test_approx_search_fallback_to_exact(vector_store: VectorStore) -> None:
 vector_store.build_index(n_clusters=2)

 query: NDArray[np.floating] = vector_store.embeddings[2]
 results = vector_store.search(
     query,
     k=2,
     use_approx=True,
     n_probe_clusters=0,
 )
 # Fails with ValueError: need at least one array to concatenate


def test_approx_search_returns_results(vector_store: VectorStore) -> None:
 vector_store.build_index(n_clusters=2)

 query: NDArray[np.floating] = vector_store.embeddings[1]
 results = vector_store.search(
     query,
     k=2,
     use_approx=True,
     n_probe_clusters=1,
 )
 assert len(results) == 2
 # Currently returns only 1 result

Expected Behavior

• Approximate search should never crash.
• Approximate search should return up to the requested number of results, falling back to exact search if needed.

Observed Behavior

• Crashes with ValueError in some cases.
• Returns fewer results than requested in other cases.

Suggested Next Steps

• Investigate why clusters can end up empty and cause concatenation errors.
• Ensure the search function gracefully falls back to exact search if there are no items in the probed clusters.
• Verify that k results are always returned whenever possible.

Notes

• This bug is only triggered in approximate search (use_approx=True) with small datasets or low n_probe_clusters.
• Exact search (use_approx=False) works correctly and all related tests pass. (See PR tests: Add initial unit tests for VectorStore #128 )