Resources
Frequently asked questions
Short answers to the questions teams ask before integrating Compresr.
Short answers to the questions teams ask before integrating Compresr. If something here is unclear, email [email protected] and we read every message.
Truncation drops the start or end of your context blindly. If the sentence that answers the user's question happens to live in the middle of a 50k-token document, naive truncation either keeps it by luck or throws it away by luck.
Compresr scores tokens by relevance to the query you pass alongside the context and drops the least relevant ones first. The evidence stays regardless of where it appears in the document. That makes the same prompt smaller without changing which question the model can actually answer. See Query-specific compression for the full mental model.
Still stuck?
Email [email protected] - we read every message and usually reply the same day.