An abstractive compressor rewrites retrieved passages into a shorter passage that preserves important information, rather than merely selecting source sentences. It reduces prompt cost and processing load while helping prevent relevant evidence from being buried in irrelevant context. section 13.0
Unlike token pruning, which eventually breaks sentences and entities, an abstractive compressor can combine evidence across documents. It is most useful when the answer depends on a relationship that no single retrieved passage states directly. section 13.1.1 section 13.4
Cross-document drug interaction
One retrieved document describes a drug’s mechanism; another describes a second medication’s contraindications. An extractive compressor can preserve both sentences, but an abstractive compressor can write a new sentence connecting the drugs to the resulting interaction. section 13.4
Choose the compression strategy according to where the answer resides:
- Use token-level pruning when documents contain self-contained answers and expendable boilerplate. section 13.0
- Use an extractive compressor when the best answer consists of sentences already present in the retrieved documents. section 13.1.1
- Use an abstractive compressor when answering requires joining facts across documents or rewriting them into a directly usable passage. section 13.4
RECOMP trains dedicated compressors offline against downstream exact-match and F1 performance rather than generic language quality. Its abstractive variant may refuse compression when retrieved evidence is unhelpful, avoiding misleading summaries. section 13.4
