Extractive summarization reduces context by deleting selected material without generating replacement text. Because it preserves original wording, it is a safer default when answers must cite source passages; abstractive compression can invent facts. section 13.9 section 13.4.1
Choose extractive pruning when the task rewards faithful retrieval of original text, then control its failure modes:
- Use it for factual recall or answers that should reproduce a span from the source; prefer abstractive synthesis for summaries or multi-hop answers combining documents. section 13.7
- Start around 4× compression, validate against uncompressed answers on held-out queries, and increase the ratio only while accuracy holds. section 13.7
- Force-keep domain-critical terms such as drug names, contractual parties, model numbers, statute identifiers, and negations. section 13.7
Moderate compression versus over-pruning
Removing distractors can improve accuracy, but beyond roughly 6× compression, meaning increasingly disappears with the deleted tokens. A well-tuned system therefore trades under-extraction against over-pruning rather than maximizing compression. section 13.9 section 13.7
Table 13.4 Method tradeoffs
|
Compression paradigm
|
Primary methodology
|
Optimal ratio
|
Compute overhead
|
Risk of semantic loss
|
|
Heuristic Pruning
|
TF-IDF Filtering
|
Low (~2x)
|
Negligible
|
Moderate
|
|
Proxy Perplexity
|
LongLLMLingua
|
Moderate (4x - 10x)
|
Moderate
|
Low to Moderate
|
|
Native Pruning
|
AttentionRAG
|
Moderate (4x - 6x)
|
Varies
|
Very Low
|
|
Trained Synthesis
|
RECOMP, CORE
|
Extreme (10x - 30x)
|
High
|
Low
|
|
KV-Cache Eviction
|
H2O, SnapKV, RazorAttention
|
N/A (decoding-time)
|
Low
|
Very Low
|
Compare heuristic filtering, proxy-perplexity methods, native attention pruning, trained synthesis, and decoding-time cache eviction by ratio, overhead, and semantic-loss risk. section 13.7




