RECOMP (Retrieve, Compress, Prepend) is a retrieval-augmented generation framework that trains compressor models offline to improve downstream task performance. Unlike generic pruning, it uses exact match and F1 signals and may refuse to add context when retrieved documents are unhelpful. section 13.4
RECOMP provides two compression strategies:
- The extractive compressor selects useful sentences from retrieved documents and stitches them together. section 13.4
- The abstractive compressor reads across documents and writes a shorter passage that may not appear verbatim in any source. section 13.4
- Selective augmentation addresses retrieval misses: instead of producing a misleading summary from irrelevant documents, the compressor can return nothing. section 13.4
Cross-document drug interaction
Suppose one retrieved document describes a drug's mechanism and another describes a second medication's contraindications. An extractive compressor can preserve relevant sentences from both; an abstractive compressor can synthesize a sentence stating the resulting interaction when neither source states that claim alone. section 13.4
RECOMP targets a limitation of token-level pruning: once sentences and entities begin to fragment, removing individual tokens cannot preserve meaning or combine facts. Its learned rewriting and refusal behavior trade additional model training for stronger compression and safer handling of irrelevant retrieval. section 13.1.1 section 13.4
