Pair Nova 2 Lite with Claude for cost-optimized document processing
TL;DR
AWS describes a Bedrock pipeline for scanned yearbook pages: Amazon Nova 2 Lite detects photos, extracts visible names with coordinates, and returns page metadata in one call. Claude Sonnet 4.6 then handles spatial matching: using the Nova JSON, the image, and page layout, it decides which name belongs to which face. In a 336-page test, the pipeline produced 3,122 name-to-face associations. AWS says 93.3 percent scored at least 0.95 confidence, while only 0.3 percent fell below 0.90.
Nauti's Take
Coming soon — Nauti's Take is being prepared.
Briefingshow
The example points to a pragmatic trend: not every document problem needs to be solved end-to-end by the strongest model. Splitting extraction, layout reasoning, and validation makes costs more controllable and lets teams replace individual stages later. For archives, catalogs, or personnel directories, that often matters more than a flashy one-prompt demo.
Sources
