HNSW Search: Changing ef Has No Effect on Recall (Milvus C++ SDK)

[AdeelAslamSoton]

Hi all,

I'm using the Milvus C++ SDK with an HNSW index and testing different ef values during search:

int total_efs[] = {20, 40, 60, 100, 200, 300, 400, 500, 600,
800, 1000, 1200, 1500, 1700, 1900, 2100};

In each search:

ef is set via AddExtraParam("ef", total_efs[i])

I'm using ParallelFor with 40 threads

TopK is 10, metric is L2

Each result is saved per query

❗ Problem
No matter how high I set ef, recall stays the same. I expected higher ef to improve recall (at the cost of speed), but that’s not happening, I already use HNSW for collection construction

❓ Questions
Does ef work at search time for HNSW in Milvus C++ SDK?

Code snippt for search is here
milvus::SearchArguments arguments;
arguments.SetCollectionName(collection_name);
arguments.SetTopK(10);
arguments.SetMetricType(milvus::MetricType::L2);
arguments.AddExtraParam("ef", total_efs[i]);

if (!data_vectors.first.empty()) {
arguments.SetExpression(data_vectors.first[row]);
}
arguments.AddTargetVector(vector_field_name, data_vectors.second[row]);
arguments.AddOutputField(filter_field_name);
arguments.AddOutputField("id");

milvus::SearchResults results;
status = client->Search(arguments, results);

Thanks 🙏

[xiaofan-luan]

what is the recall you get?

can you test mulitple different queries and see the result of queries?

adding ef to too high is not helping, for topk=10 usually ef 200 is good. enough.

the other way to improve recall is at index build size, you can do

milvus::IndexParam index_param("vector_field_name", "HNSW");
index_param.AddExtraParam("M", 64); // controls graph degree
index_param.AddExtraParam("efConstruction", 200); // affects recall
client->CreateIndex(collection_name, index_param);

add see if it improves the recall

[AdeelAslamSoton]

what is the recall you get?

can you test mulitple different queries and see the result of queries?

adding ef to too high is not helping, for topk=10 usually ef 200 is good. enough.

the other way to improve recall is at index build size, you can do

milvus::IndexParam index_param("vector_field_name", "HNSW"); index_param.AddExtraParam("M", 64); // controls graph degree index_param.AddExtraParam("efConstruction", 200); // affects recall client->CreateIndex(collection_name, index_param);

add see if it improves the recall

Thanks for your response!

Just to clarify, I am changing the queries during the experiments. For the HNSW index construction, I used the following parameters:

M = 64

efConstruction = 250

However, I'm consistently getting a recall of 0.5, regardless of the ef value used during search. I was expecting the recall to increase with higher ef values, but that’s not happening.

[xiaofan-luan]

what is the recall you get?
can you test mulitple different queries and see the result of queries?
adding ef to too high is not helping, for topk=10 usually ef 200 is good. enough.
the other way to improve recall is at index build size, you can do
milvus::IndexParam index_param("vector_field_name", "HNSW"); index_param.AddExtraParam("M", 64); // controls graph degree index_param.AddExtraParam("efConstruction", 200); // affects recall client->CreateIndex(collection_name, index_param);
add see if it improves the recall

Thanks for your response!

Just to clarify, I am changing the queries during the experiments. For the HNSW index construction, I used the following parameters:

M = 64

efConstruction = 250

However, I'm consistently getting a recall of 0.5, regardless of the ef value used during search. I was expecting the recall to increase with higher ef values, but that’s not happening.

I think it doesn't make sense.

It's better to check your ground truth and the result you get. it could be a bug or due to wrong ground truth.

I do know several corner cases that might cause low recalls, you should also try multiple quries and see if you still get same recalls