An extractive compressor selects useful sentences from retrieved documents and stitches them together without generating new wording. It reduces context before a downstream generator consumes it. section 13.1.1 section 13.4
RECOMP trains the compressor offline against downstream task performance, using exact match and F1 rather than perplexity or generic quality measures. The compressor may also refuse when the retrieved documents contain nothing useful, avoiding misleading prompts. section 13.4
Use the extractive approach when:
- Each retrieved document already contains a self-contained answer, so preserving the strongest sentences is sufficient. section 13.4
- The answer requires combining facts that no individual source states, prefer an abstractive compressor instead. section 13.4
- Avoid treating compression as automatically reliable: misleading compressed prompts can be worse than injecting nothing. section 13.4
Cross-document drug interaction
Suppose retrieval returns one document describing a drug's mechanism and another describing a second medication's contraindications. An extractive compressor can retain relevant sentences from both, but cannot state the interaction if neither source states it directly; an abstractive compressor can synthesize that combined claim. section 13.4
