Topics

  • Single embeddings generated from entire text struggle to represent diverse information in long documents
  • Mathematical limitation: cosine similarity between query and document is average similarity to document tokensquestion
  • Adding irrelevant information to a document reduces relevance for all queries