Compresr

Teams sending long documents or prompts to LLMs pay for far more tokens than the model actually needs to answer a question. Common workarounds like trimming or truncation cut off the tail of a document, often removing the exact passage that held the answer, causing accuracy to collapse on long documents. Summarization is another common approach, but it produces a lossy rewrite that loses nuance and exact wording, while adding extra LLM calls and latency. Question-agnostic compression tools compress blindly without knowing the query, so they keep irrelevant tokens and drop important ones, rarely achieving more than 5x compression without hurting accuracy.

Compresr provides a question-aware compression API: you feed it both the query and the context, and it returns only the tokens that actually affect the answer. On a Boeing 10-K example shown on the site, this took a 112,552 token document down to 498 tokens, a 226x compression that cut per-query cost from $0.263 to $0.037 and reduced latency from 18s to 13.7s. On an independent FinanceBench benchmark (128 questions over SEC filings), light compression (~2x) with GPT-5.2 held accuracy at 77% versus 73% for the uncompressed baseline while cutting context from ~106K to ~56K tokens and saving about 47% on cost. The product ships as a TypeScript/Python SDK with a hosted option (one API key, usage dashboard, volume pricing) or an on-prem deployment that runs inside a customer's VPC for regulated workloads like finance and healthcare.