Question about Vector Search Behavior with and without Radius in Milvus

[dmitriivahrushev]

Dear Milvus Team,

First of all, thank you for the great work you are doing on Milvus. We really appreciate the platform and the flexibility it provides for vector search.

I am trying to better understand how vector search behaves internally in two slightly different scenarios, and I would be very grateful if you could help clarify this.

We are using COSINE metric and experimenting with two cases:

Case 1 – Using limit and radius:
def batch_search(self, collection_name, data, limit, radius):
    search_params = {"metricType": "COSINE", "params": {"radius": radius}}
    result = self.client.search(
        collection_name=collection_name,
        data=data,
        limit=limit,
        search_params=search_params,
    )
    return result

Case 2 – Using limit only (no radius):
def batch_search(self, collection_name, data, limit):
    search_params = {"metricType": "COSINE"}
    result = self.client.search(
        collection_name=collection_name,
        data=data,
        limit=limit,
        search_params=search_params,
    )
    return result

What we observe is that the results differ significantly between these two cases — even when the radius value is set to a very small threshold. It feels like the search behavior itself changes.
Could you please help clarify the following:
-When radius is not specified, is this still a pure Top-K search?
-Is radius simply an additional post-filter after Top-K retrieval?
-Or does adding radius actually change the internal search logic or candidate selection strategy?
-Are there any algorithm-level differences between these two scenarios (especially when using HNSW indexes)?

We would really appreciate a more detailed explanation of how these two modes differ internally, as this impacts how we tune our search thresholds and ranking logic.
Thank you very much for your time and support!

[xiaofan-luan]

Thank you for your question.

When radius is not specified, the search behaves as a standard ANN Top-K search.

Range search and Top-K search are fundamentally different. They follow different execution paths internally, so it is expected that results may differ even when the radius is set to a very small value.

Range search prioritizes distance ordering (nearest to farthest) within the specified threshold. To guarantee this behavior, the algorithm may skip certain candidates during traversal. As a result, its behavior is not equivalent to applying a simple post-filter on Top-K results.

Additionally, please note that Zilliz Cloud includes index-level optimizations beyond the open-source Milvus defaults. In practice, you can often reduce discrepancies between Top-K and range search results by increasing search parameters (e.g., ef, nprobe, etc.), which improves recall.

[dmitriivahrushev]

Massive shoutout for the lightning-fast help today!