liveBook · Manning

Retrieval-extractive pipeline

As explained in Retrieval Augmented Generation, The Foundational Ideas

A retrieval-extractive pipeline retrieves relevant passages and then extracts a literal answer span from each passage.

A retrieval-extractive pipeline answers open-domain questions in two stages: a sparse retriever finds relevant passages, then a comprehension model extracts a literal answer span from each passage. This architecture makes large-scale extractive question answering reproducible, but it struggles when the answer requires synthesizing facts scattered across documents. section 5.1

A typical implementation separates offline corpus access from online question answering:

  • Search a large corpus such as Wikipedia with TF-IDF, BM25, or bigram hashing. section 5.1
  • Pass each retrieved passage independently to a reader, originally RNN-based and later BERT-based. section 5.1
  • Return the literal text span identified as the answer rather than generating a synthesized response. section 5.1

DrQA

DrQA combines bigram-hashing and TF-IDF retrieval over Wikipedia with a multilayer neural reader that locates answer spans inside retrieved paragraphs. section 5.1

Figure 5.1 DrQA Retrieval Extractive Pipeline.

Use the pipeline diagram to trace retrieval, passage reading, and answer-span extraction as separate stages. section 5.1

How Retrieval Augmented Generation, The Foundational Ideas covers it

How deep it goes
Explains retrieval methods, reader evolution, and answer-span extraction, but stops short of multi-document evidence fusion.
Examples
Uses DrQA over Wikipedia as the canonical concrete implementation.
What you can do after
Build it from scratch as a retriever followed by a passage-level span reader.
How it connects
Connects sparse retrieval with TF-IDF, BM25, bigram hashing, BERT, DrQA, RAG, and Fusion-in-Decoder.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Retrieval Augmented Generation, The Foundational Ideas. The section links point to the parts of the book it draws on.