Skip to content
GenAI IDP

Granular Assessment Retirement

Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved. SPDX-License-Identifier: MIT-0

Granular Assessment Retirement

Section titled “Granular Assessment Retirement”

This note documents the retirement of granular assessment and what it means for your configuration. The short version: no customer action is required — existing configs continue to validate and run unchanged.

What changed

Section titled “What changed”

Granular assessment is retired and its service has been deleted. Granular assessment was a separate service that fanned per-attribute / per-list-batch confidence tasks across a thread pool with DynamoDB caching. It was the historical mechanism for scoring large lists on the non-agentic path, but:

  • its <<CACHEPOINT>> caching was ineffective (a concurrent cacheWrite storm), and
  • it cost roughly 4–5× more than a consolidated pass for equal accuracy.

In A/B testing on a 120-row bank statement, turning granular off yielded −78% Bedrock cost at equal accuracy, with full per-cell coverage plus geometry — whereas granular actually produced 0% geometry.

Its capability is now fully covered by standalone large-list batching (see below).

granular.* config keys are now a no-op

Section titled “granular.* config keys are now a no-op”

Any extraction.confidence.granular.* (or legacy assessment.granular.*) keys in an existing configuration continue to validate but are ignored. The confidence config model uses extra="ignore", so stale granular keys neither raise nor take effect.

No config edit is required. You do not need to remove granular.* keys. They are silently dropped at runtime.

The single knob for large-list assessment: list_batch_size

Section titled “The single knob for large-list assessment: list_batch_size”

Large lists (bank statements with hundreds of transactions, line-item tables, brokerage holdings) are now assessed by standalone large-list batching. The standalone Assessment step slices the largest list field into extraction.confidence.list_batch_size chunks (default 25), assesses each chunk sequentially with the shared scalars/context, and reconciles the results so every list cell gets its own confidence and bounding box — automatically, with no additional configuration.

extraction:
  confidence:
    enabled: true
    mode: separate            # off | separate (default) | integrated
    list_batch_size: 25       # rows per assessment batch for large lists
  • Lower list_batch_size if a model still struggles to enumerate a full chunk.
  • Raise it to reduce the number of inference calls.

See Large-list batching in the Extraction & Confidence guide.

Mode guidance for large documents

Section titled “Mode guidance for large documents”
  • Advanced (agentic) extraction is recommended for large documents and big tables. It shards both extraction and confidence assessment, and produced the best-calibrated confidence in A/B testing. For 100+ page documents prefer runtime: step_functions and raise max_concurrent_batches.
  • Simple + separate confidence remains fully viable for large lists via list_batch_size batching — full per-cell confidence and geometry with no extra configuration. This is the direct replacement for granular assessment.
  • Rule of thumb: a document that is large only because it contains a long list is well served by Simple + separate + batching; a document with a very large single section is better served by Advanced sharding.

See Large-Document Guidance.

Configuration format is unchanged (stays v0.6)

Section titled “Configuration format is unchanged (stays v0.6)”

This retirement does not introduce a new config version. Configurations remain v0.6: confidence lives under extraction.confidence.*, geometry under extraction.geometry.*, and HITL under the top-level hitl.*. Pre-v0.6 configs are still migrated automatically on read.

Section titled “Related Documentation”