Natural Language Processing · 免费预览
稠密检索与 RAG
词表不匹配问题；采用最近邻搜索的双编码器稠密检索；基于可信段落的检索增强生成。
1. Dense Retrieval and RAG
Dense Retrieval and RAG
2. Bi-encoder score
A query embedding is and a document embedding is . Compute their dot-product relevance score, rounding to three decimals.
3. Why sparse retrieval missed
A query asks for something to eat and the best passage is a chowder recipe, with no query word appearing in it. What is this failure called, and which fix addresses it?
4. Choosing the encoder arrangement
Why do large-scale systems use a bi-encoder for first-pass retrieval and a cross-encoder only for reranking the top candidates?
5. What retrieval adds over parameters
Which set of problems with prompting a pretrained model does retrieval-augmented generation directly target?
6. Beyond the basic RAG loop
A question needs two facts that live in different documents, and the retrieved set also contains an irrelevant passage. Which two extensions address these issues?