Prompt Compression

Long prompts may contain repeated instructions, verbose context, or retrieved passages that are not useful for a particular answer. Prompt compression tries to shorten this input before sending it to a model. It can be manual, such as removing redundancy, or algorithmic, such as selecting tokens or rewriting content into a shorter representation. Research systems such as LLMLingua and LLMLingua-2 study prompt compression for reducing inference cost and latency while attempting to preserve task performance. The results apply to the paper’s models, datasets, and compression settings; they do not guarantee that every prompt or application will remain unchanged after compression. Compression has tradeoffs. A short phrase may be essential to a constraint, exception, citation, or safety requirement. A compression model may omit low-frequency details that matter to a user. Algorithmic compression also takes compute and can increase total latency if the prompt is short or the compressor is slow. Token savings reduce cost only when provider pricing and cache behavior make those tokens billable or operationally relevant. Start by removing repeated boilerplate and keeping task goals, constraints, definitions, examples, and relevant evidence intact. If using a compressor, compare original and compressed prompts on representative test cases. Measure input tokens, compression time, response latency, cost, and quality. Treat compression as an optimization to validate, not as a guarantee that fewer tokens preserve meaning.