跳到主要内容Skip to main content

混合检索Hybrid retrieval

Hybrid retrieval

  • 数据与检索Data & retrieval
  • 第三站 · 用深Station 3 · Going Deeper
  • 引用核验 2026-11-19refs checked 2026-11-19
解决什么问题What problem it solves

把关键词检索(BM25)和向量检索的结果融合,兼顾精确词与语义近义。Fuse keyword retrieval (BM25) with vector retrieval so exact terms and paraphrases both work.

这是什么What it is

同一查询分别走稀疏检索和稠密检索,各拿一组候选。

用 RRF 或加权分数把两组结果合并、去重、重排。

精确编号、专有名词靠关键词;改写、同义表达靠向量。

The same query runs through sparse and dense retrieval, producing two candidate sets.

RRF or weighted scores merge, deduplicate and rerank the two sets.

IDs and proper nouns rely on keywords; paraphrases rely on vectors.

为什么需要知道Why it matters

只做向量检索会漏掉产品型号、错误码、合同编号这类「必须字面命中」的内容。Vector-only retrieval misses product models, error codes and contract numbers that must match literally.

出处Sources
  • Elastic 混合检索文档Elasticsearch hybrid search documentation ↗