检索--混合检索

专题4：混合检索

检索代码如下所示：

from pymilvus import AnnSearchRequest
#### STEP1：构造检索模块request_1 和request_2 
## 假设稠密数据
query_dense_vector = [0.3580376395471989, -0.6023495712049978, 0.18414012509913835, -0.26286205330961354, 0.9029438446296592]
search_param_1 = {
    "data": [query_dense_vector],
    "anns_field": "dense",
    "param": {
        "metric_type": "IP",
        "params": {"nprobe": 10}
    },
    "limit": 2
}
request_1 = AnnSearchRequest(**search_param_1)
## 假设稀疏数据
query_sparse_vector = {3573: 0.34701499565746674}, {5263: 0.2639375518635271}
search_param_2 = {
    "data": [query_sparse_vector],
    "anns_field": "sparse",
    "param": {
        "metric_type": "IP",
        "params": {}
    },
    "limit": 2
}
request_2 = AnnSearchRequest(**search_param_2)
reqs = [request_1, request_2]
#### STEP2：利用检索模块融合进混合检索
from pymilvus import MilvusClient
res = client.hybrid_search(
    collection_name="hybrid_search_collection",
    reqs=reqs,
    ranker=ranker,
    limit=2
)
for hits in res:
    print("TopK results:")
    for hit in hits:
        print(hit)

Rerankers策略主要有两种方法如下：

方法1：rrf加权策略

文档d在所有检索器的排名倒数和

from pymilvus import RRFRanker
ranker = RRFRanker(100)

方法2：对不同的检索空间的document先取交集，然后对score进行加权策略，策略如下：

图片来自于最佳实践论文

from pymilvus import WeightedRanker
rerank= WeightedRanker(0.8, 0.3)