Class_note3.md

RAG

Definition

• retrieval(检索) augmented generation: combine retrieval and generation, Leveraging external knowledge bases to enhance LLM performance

Principle

• knowledge source -> vector-DB query -> retrieval -> vector-DB retrieval -> generation

Optimization

• enhance the quality of DB
  • embedding optimization: multi-task
  • indexing: meta data
• query
• context: rerank.eg
• retrieval
  • iterative retrieval
  • recursive retrieval: chain of thought
  • adaptive retriveval: Flare, self-RAG
• LLM fine-tuning

RAG vs. Fine-tuning

• RAG

  

• Fine-tuning

  

Metrics

acc, recall, f1 score, bleu, rouge tool: RAGAS, ARES, TruLens

Huixiangdou

• 智能客服：技术支持、领域知识对话
  • 群组中存在闲聊
  • parse users' intent
• protocol: BSD-3-Clasue
• workflow