Retrieve, read, and generate

Retrieve-read architectures differ mainly in where retrieved evidence interacts with the reader or generator.

Retrieve, read, and generate: architecture comparisonEach lane reads from input to output.

Retrieve then read

  1. retrievePassage search
  2. readSpan or answer reader
  3. outputAnswer
  • Memory Networks
  • End-To-End Memory Networks
  • DrQA
  • ORQA

Latent-document RAG

  1. retrieveTop-k documents
  2. conditionDocument likelihood
  3. marginalizeSequence or token
  4. outputGenerated answer
  • RAG-Sequence
  • RAG-Token

Fusion-in-Decoder

  1. retrieveMany passages
  2. encodeIndependent encoders
  3. fuseDecoder attention
  4. outputGenerated answer
  • FiD
  • FiD-KD
  • KG-FiD

Retriever learning

  1. observeReader utility
  2. distillPassage targets
  3. trainRetriever update
  4. evaluateEnd-to-end result
  • EMDR²

Read the full explanation