[Bug]: mh_lsh_band parameter lacks validation - accepts invalid values (0, -1, >num_hashes)

[zhuwenxing]

Is there an existing issue for this?

• I have searched the existing issues

Environment

• Milvus version: master
• Deployment mode: standalone
• MQ type: rocksmq
• SDK version: pymilvus 2.7.0rc136
• OS: Linux (K8s deployment)

Current Behavior

When creating a MINHASH_LSH index, the mh_lsh_band parameter accepts clearly invalid values without any error or warning:

1. mh_lsh_band=0: Server accepts it, index is created, and search returns results (likely using a fallback/default)
2. mh_lsh_band=-1: Server accepts it, index is created, and search returns results
3. mh_lsh_band > num_hashes (e.g., band=26 with num_hashes=16): Server accepts it, index is created, and search returns results

All three cases should be rejected with clear validation errors.

Expected Behavior

The server should validate mh_lsh_band at index creation time:

• mh_lsh_band must be a positive integer (> 0)
• mh_lsh_band must not exceed num_hashes (from the MinHash function params)
• mh_lsh_band should ideally be a divisor of num_hashes for optimal LSH behavior

Invalid values should be rejected with a clear error message at index creation time.

Steps To Reproduce

from pymilvus import MilvusClient, DataType, Function, FunctionType

client = MilvusClient(uri="http://<host>:19530")

schema = client.create_schema(enable_dynamic_field=False)
schema.add_field("id", DataType.INT64, is_primary=True, auto_id=False)
schema.add_field("text", DataType.VARCHAR, max_length=65535)
schema.add_field("minhash_sig", DataType.BINARY_VECTOR, dim=512)

schema.add_function(Function(
    name="text_to_minhash",
    function_type=FunctionType.MINHASH,
    input_field_names=["text"],
    output_field_names=["minhash_sig"],
    params={"num_hashes": 16, "shingle_size": 3},
))

# Case 1: mh_lsh_band = 0
index_params = client.prepare_index_params()
index_params.add_index(
    field_name="minhash_sig",
    index_type="MINHASH_LSH",
    metric_type="MHJACCARD",
    params={"mh_lsh_band": 0},  # Should be rejected
)
client.create_collection("test_band_0", schema=schema, index_params=index_params)
# No error! Index created successfully.

# Case 2: mh_lsh_band = -1
index_params2 = client.prepare_index_params()
index_params2.add_index(
    field_name="minhash_sig",
    index_type="MINHASH_LSH",
    metric_type="MHJACCARD",
    params={"mh_lsh_band": -1},  # Should be rejected
)
client.create_collection("test_band_neg", schema=schema, index_params=index_params2)
# No error! Index created successfully.

# Case 3: mh_lsh_band = 26 > num_hashes = 16
index_params3 = client.prepare_index_params()
index_params3.add_index(
    field_name="minhash_sig",
    index_type="MINHASH_LSH",
    metric_type="MHJACCARD",
    params={"mh_lsh_band": 26},  # Should be rejected (> num_hashes=16)
)
client.create_collection("test_band_exceed", schema=schema, index_params=index_params3)
# No error! Index created successfully.

Anything else?

Root cause analysis:

There is no validation for mh_lsh_band at any layer:

1. Proxy layer: No parameter range check when processing CreateIndex request
2. Index node: No validation when building the MINHASH_LSH index
3. Knowhere: The underlying index library silently accepts invalid band values

The fix should add validation at the proxy layer (index creation path) or in the Knowhere MINHASH_LSH index config to ensure:

• mh_lsh_band > 0
• mh_lsh_band <= num_hashes (requires cross-referencing the collection schema's MinHash function params)
• Optionally warn if num_hashes % mh_lsh_band != 0 (non-divisor bands may cause suboptimal LSH behavior)