Optimization guide
Compression ratio vs answer quality
Higher compression saves more money but may reduce answer quality. The key is finding the compression budget where savings are maximized without noticeable quality loss.
Empirical findings
SuperCompress benchmarks show:
- Up to 65% compression: No measurable quality loss (100% oracle recall)
- 65-80% compression: Minor quality impact (95-99% oracle recall)
- 80-90% compression: Moderate quality impact (85-95% oracle recall)
- 90%+ compression: Significant quality risk (below 85% oracle recall)
Frequently asked questions
Does the optimal ratio vary by use case?
Yes. Factual extraction tasks (contract review, data extraction) need lower ratios. Creative tasks (content generation, brainstorming) can tolerate higher ratios.
How do I test for my use case?
Run 50 examples at each compression level and compare outputs manually or with an evaluation LLM.
Build with less context
Put compression in front of your next LLM call.
Use the hosted API or run SuperCompress locally. Keep the evidence, drop the token waste, and measure the savings before it reaches your model.