#AI

Embedding Models: from Architecture to Implementation

3 Contextualized token embeddings

单词嵌入:word2vec/glove无法捕获上下文,BERT可以捕获上下文

BERT(an encoder only transformer model)被大量用作 sentence embedding model 的组件

ref

4 Token vs. sentence embedding

dual encoder 架构做 Q/A时的 sentence embedding

5 Training a dual encoder

用对比学习训练

6 Using embeddings in RAG

7

用embeddings快速获取文档,再用更准确的Cross-encoder重排序


Understanding and Applying Text Embeddings

6

先Top K,再Top P,再Temperature

参考