Milvus Vector ANN Search

[dmitriivahrushev]

Hello wonderful Milvus team! 👋
We hope this message finds you well! We're absolutely thrilled about your amazing product and could really use your expert guidance on our journey.
We're currently in the testing phase, exploring the possibility of adopting Milvus (which we're super excited about!) while comparing it with our current OpenSearch setup.

Our current situation:
We have OpenSearch up and running with HNSW index (metric type: COSINE) using m16_ef1024 parameters and performing KNN search:

{
    "size": "25",
    "query": {
        "knn": {
            "my_vector": {
                "k": 1024
            }
        }
    }
}

Our Milvus setup:
We've created what we believe is an equivalent HNSW index with the same parameters (m16_ef1024) and performing search:

{
    "limit": 25,
    "searchParams": {
        "metricType": "COSINE",
        "params": {
            "ef": 1024
        }
    }
}

We're getting about 97% recall.

Our friendly questions for you:
-Is it possible to perform exact KNN search (rather than ANN) with HNSW index in Milvus?
-How valid/appropriate is it to compare OpenSearch's KNN results with Milvus's ANN results?

We'd be incredibly grateful for any insights, tips, or suggestions you might have. We're really passionate about making the right choice and would love to join the Milvus family!
Thanks a million for your time and expertise!

[xiaofan-luan]

Q1: Can Milvus do exact KNN search with HNSW index?

No, HNSW is an ANN (approximate nearest neighbor) algorithm by design, so exact KNN is not supported with HNSW. However, you can absolutely push recall beyond 97% — achieving 99% or even 99.5% is
realistic by tuning your search parameters (e.g. increasing ef). The 97% you're seeing likely has room for improvement depending on your query parameters and data distribution.

Q2: Is it fair to compare OpenSearch KNN results with Milvus ANN results?

Not really — this isn't an apples-to-apples comparison. KNN (brute-force exact search) is significantly slower than ANN, so comparing them directly isn't fair to either system. The right way to
benchmark is to compare both under the same search mode (ANN vs ANN) with equivalent parameters, and measure recall against a known ground truth (e.g. brute-force results from a FLAT index). That way
you get a meaningful comparison of both accuracy and performance.

[xiaofan-luan]

you can do KNN with FLAT index

[dmitriivahrushev]

Thanks so much for your clear and helpful response! It really clarified things for me. I appreciate you taking the time to explain it so well!