L Li Peng
2 min read
Contextual Retrieval: Empirical Study

3 phases · 6 controlled experiments · reproducible

  • Python
  • RAG
  • Contextual Embeddings
  • BM25
  • Knowledge Graph

Reproduced Anthropic’s Contextual Retrieval (contextual embeddings + contextual BM25), then designed systematic comparative experiments across structured and government-document datasets. This work is the basis of my first empirical research paper.

Experiments

PhaseDatasetContent
1Canteen menu (structured lists)Baseline vs CR vs Jieba+KG
2Flood-prevention plans (gov docs)Baseline vs CR vs Deep KG
3Same, OneKE-13B + OpenKGKnowledge-graph schema optimization

Core finding: CR is a double-edged sword across data types; knowledge graphs still hit real bottlenecks under current conditions.