milvus auto index

[rere950303]

• milvus version : v2.4.23
• mode : cluster

---

• There is an indexing type called AutoIndex for vector or scalar fields. Compared to building an index manually by specifying parameters for HNSW, DiskANN, etc., which approach is more recommended—manual indexing or using AutoIndex?

• In the case of Zilliz Cloud, AutoIndex seems to automatically optimize indexing and querying using machine learning techniques. Is the same mechanism also applied in Milvus?

[yhmo]

For open-source milvus, the AUTOINDEX default parameters can be configured in the default milvus.yaml.
The relevant code shows that for FloatVector, AUTOINDEX is actually HNSW, and the parameters are configured by autoIndex.params.build:
milvus/pkg/util/paramtable/autoindex_param.go

 DefaultValue: `{"M": 18,"efConstruction": 240,"index_type": "HNSW", "metric_type": "COSINE"}`, 

So, if you want to change the parameters of AUTOINDEX, you can add these lines at the end of milvus.yaml:

autoIndex:
  params:
    build: {"nlist": 240,"index_type": "IVF_FLAT", "metric_type": "COSINE"}

In most of cases, no need to reset its parameters.

For Zilliz Cloud, the AUTOINDEX setting is not publicly available. The index used for AUTOINDEX is also different from open-source Milvus.

[rere950303]

oh i see..
thanks for answering!! 👍