LLM context optimization

Long prompts raise input cost, increase latency and can bury the evidence needed for a focused answer. Context optimization removes lower-priority material for the current task while keeping names, dates, numbers, conditions and instructions available.

A summary rewrites information. Evidence-aware selection retains source passages and their meaning. This distinction matters when an exception, threshold or negation changes the answer.

Keep source context separate from the task, extract answer and evidence anchors from the query, build a smaller candidate, verify protected details, then return the original when the candidate cannot pass the gates.

Track effective token saving, critical-evidence recall, linked-evidence preservation, answer equivalence, unsafe accepted candidates, passthrough rate, recovery cost and p95 latency.

TrimLayer returns optimized_context for source text and an optional query. It does not make the next model call: your application keeps control of the destination, provider and response validation.