AWS proposes task-aware knowledge compression at 1.6% of baseline token cost.

The AWS ML blog describes pre-compressing a knowledge base into task-specific representations instead of retrieving chunks at query time, so different tasks get different compressions of the same source. Four tiers from 8x (~87.5% context reduction) to 64x (~98.4%). A 100K-token knowledge base at 1,000 queries/day costs 6.25%, 3.1%, and 1.6% of baseline versus 10% for top-10-chunk RAG. Reference architecture is two serverless pipelines on Lambda, Bedrock, ElastiCache Serverless, S3, API Gateway, and Cognito.