-
Notifications
You must be signed in to change notification settings - Fork 302
向量索引配置
xuxijie edited this page Dec 9, 2022
·
3 revisions
{
"attributes": [
"id"
],
"summarys": {
"compress": true,
"summary_fields": [
"id"
]
},
"fields": [
{
"field_name": "id",
"field_type": "UINT32"
},
{
"field_name": "vector",
"field_type": "STRING"
},
{
"field_name": "DUP_id",
"field_type": "RAW"
},
{
"field_name": "DUP_vector",
"field_type": "RAW"
}
],
"index":[
{
"index_fields":[
{
"boost":1,
"field_name":"DUP_id"
},
{
"boost":1,
"field_name":"DUP_vector"
}
],
"index_name":"vector_index",
"index_type":"CUSTOMIZED",
"indexer":"aitheta_indexer",
"parameters":{
"use_linear_threshold":"10000",
"build_metric_type":"l2",
"search_metric_type":"ip",
"use_dynamic_params":"1",
"dimension":"128",
"index_type":"hc"
}
}
]
}
引入分类的目的是为了支持按照分类进行向量检索,比如一个图片有不同的类别,如果不指定分类构建向量索引,只是对检索出来的向量进行过滤很可能会出现无结果的情况。查询时需要指定待检索的类目,如果不指定的话引擎会查询所有类目,性能会急剧下降。
{
"attributes": [
"id",
"category"
],
"summarys": {
"compress": true,
"summary_fields": [
"id",
"category"
]
},
"fields": [
{
"field_name": "id",
"field_type": "UINT32"
},
{
"field_name":"category",
"field_type":"UINT32"
},
{
"field_name": "vector",
"field_type": "STRING"
},
{
"field_name": "DUP_id",
"field_type": "RAW"
},
{
"field_name": "DUP_category",
"field_type": "RAW"
},
{
"field_name": "DUP_vector",
"field_type": "RAW"
}
],
"index":[
{
"index_fields":[
{
"boost":1,
"field_name":"DUP_id"
},
{
"boost": 1,
"field_name": "DUP_category"
},
{
"boost":1,
"field_name":"DUP_vector"
}
],
"index_name":"vector_index",
"index_type":"CUSTOMIZED",
"indexer":"aitheta_indexer",
"parameters":{
"use_linear_threshold":"10000",
"build_metric_type":"l2",
"search_metric_type":"ip",
"use_dynamic_params":"1",
"dimension":"128",
"index_type":"hc"
}
}
]
}
- field_name:构建向量索引的字段,必须为RAW类型,最少包括2个字段,一个是主键(或者是主键的hash值),字段值必须是整数,另一个是包含向量的字段。如果需要对向量按照分类构建索引,可以新增一个分类字段,字段类型为RAW类型,字段值为整数。这些字段在索引中的顺序必须和在fields配置的顺序相同,而且必须是按照主键、类目(如果有)、向量这个顺序。
- index_name:向量索引的名称。
- index_type:索引类型,必须为CUSTOMIZED。
- indexer:构建向量索引的插件,目前仅支持aitheta_indexer。
- parameters:向量索引的构建参数和查询参数,不同索引类型参数不同,详见向量索引构建参数。