Document

opensearch 경험 공유

1 min read

Background 설명

RAG 시스템에서 왜 VectorDB로 OpenSearch를 사용하는가?

vector search(semantic) 와 full-text search(lexical) 둘 다 가능하기에,

검색에 대한 정확도를 올리기 위해 해당 과정 필요 https://ieeexplore.ieee.org/abstract/document/10707868

Cluster 고가용성 필요

  • vector

임베딩 모델

  • BAAI/bge-m3
  • BAAI/bge-reranker-v2-m3

Embedding flow

  1. get document from storage
  2. get embedding vector with model
  3. insert document with vector

Retrieval flow

  1. user query embedding
  2. vector search from OpenSearch
  3. user query full-text search from OpenSearch
  4. combine two result sets
  5. reranking result sets (with reranker model)
  6. return similar documents

OpenSearch KNN indexing

실험 1

일반 index (not knn)

Search

knn index

cold start

after warm-up

실험2 dataset 크기

실험3 disk latency

400K

600K

800k

1M

실험3

disk를 nfs로 변경하여 실험

그래프 뷰