AWS: Nova 2 Lite and Claude — Cost-Optimized Pipeline for Processing Scanned Documents

AWS published a two-stage pipeline architecture for mass digitization of scanned documents on the Amazon Bedrock platform. The demonstration uses school yearbooks with complex layouts, but is applicable to any large-volume document digitization scenario where visual elements need to be linked to textual data.

How the Two-Stage Pipeline Works

The two-stage pipeline is an architectural pattern in which two models work sequentially — the cheaper model handles routine processing, while the more expensive model takes over only the most demanding steps. In this case, Amazon Nova 2 Lite goes through each scanned page in a single API call: it detects photos with bounding boxes, extracts visible names with positions on the page, and emits page-level metadata. Nova 2 Lite operates at a LOW reasoning level to keep costs as predictable as possible — with a fixed price per image, regardless of resolution.

Division of Work: Nova 2 Lite vs Claude

Claude Sonnet 4.6 enters in the second step and takes on what Nova 2 Lite didn’t solve: spatial reasoning to associate names with specific faces on the page. Claude uses adaptive thinking and adjusts the depth of reasoning according to the complexity of the input — without the need for manual prompt engineering for each layout type. Nova 2 Lite generates approximately 1,000 output tokens per page; Claude consumes approximately 1,700 tokens.

What Does the Cost-Tiered Approach Deliver?

The entire point is that Claude is used exclusively where necessary. Result: the two-stage pipeline costs approximately $0.033 per page, while processing with Claude alone would cost ~$0.10. On 100,000 pages the difference is approximately $6,500. AWS tests accuracy on a sample of 3,122 name-face associations across 336 yearbook pages — 93.3% receive a confidence score ≥ 0.95, and only 0.3% fall below 0.90.

Where Else Does This Pattern Apply?

Beyond yearbooks, the same architectural pattern applies to historical archives, staff directories, real estate registries, and product catalogs. Any scenario with a clear “light” and “heavy” part of document processing is a candidate for a cost-tiered multi-model pipeline: let Amazon Nova 2 Lite handle the routine, and leave the spatially and semantically demanding steps to Claude.