v0.5 → v0.6 Migration (Assessment merged into Extraction)

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

v0.5 → v0.6 Migration (Assessment merged into Extraction)

This note documents the v0.6 configuration change that retired the standalone top-level assessment: block and folded its settings into extraction.confidence, extraction.geometry, and a top-level hitl block. The short version: no customer action is required — pre-v0.6 configurations are migrated automatically on read/update, and the migration is idempotent.

What changed

In v0.5.x, per-field confidence was configured as if it were a separate processing stage under a top-level assessment: block. v0.6 reframes confidence as an optional output of extraction, so every knob that used to live under assessment now has a new home:

• Confidence → extraction.confidence.*
• Geometry (bounding boxes) → extraction.geometry.*
• Human review (HITL) → top-level hitl.*

The single on/off/mode control is now extraction.confidence.mode, which takes one of three values:

confidence.mode	Meaning
off	No confidence scoring at all — no extra model pass, no explainability_info.
separate (default)	Confidence scored in a distinct inference (a per-shard second pass for agentic extraction, or the standalone Assessment step for simple extraction).
integrated	The extraction inference emits each value’s confidence in one pass, saving a model call (the standalone Assessment step is bypassed).

extraction.confidence.enabled still exists but is derived from mode (enabled = mode != 'off'); do not set it directly — use mode.

”Assessment disabled” → confidence.mode: off

If your v0.5.x configuration had Assessment turned off, it was stored as:

# v0.5.x
assessment:
  enabled: false

On upgrade this is migrated to:

# v0.6
extraction:
  confidence:
    mode: 'off'

A falsy assessment.enabled (false, or the strings "false", "0", "no", "off") is treated as authoritative and forces confidence.mode: off, so a disabled config stays disabled after the upgrade. This holds even for a sparse stored delta (assessment.enabled: false alone), because the migration runs on your custom delta before it is merged onto the v0.6 defaults (whose default mode is separate).

Full setting mapping (v0.5.x → v0.6)

v0.5.x (old)	v0.6 (new)	Notes
assessment.enabled: false	extraction.confidence.mode: off	Disable → off.
extraction.assessment_integration	extraction.confidence.mode	separate / integrated folded into the single mode knob.
assessment.model	extraction.confidence.model	Passthrough. LambdaHook supported.
assessment.model_lambda_hook_arn	extraction.confidence.model_lambda_hook_arn	Passthrough.
assessment.system_prompt	extraction.confidence.system_prompt	Passthrough.
assessment.task_prompt	extraction.confidence.task_prompt	Confidence-only task prompt.
assessment.temperature / top_p / top_k / reasoning_effort	extraction.confidence.*	Passthrough (same names).
assessment.image.*	extraction.confidence.image.*	Passthrough.
assessment.inshard_list_batch_size	extraction.confidence.list_batch_size	Renamed. Default 25.
assessment.geometry_mode: llm_with_ocr_grounding	extraction.geometry.mode: llm_grounded	Value renamed.
assessment.geometry_mode: llm_only	extraction.geometry.mode: llm	Value renamed.
assessment.geometry_mode: ocr_only	extraction.geometry.mode: ocr_only	Unchanged.
assessment.ground_geometry_in_ocr: false	extraction.geometry.mode: llm	Legacy flag → use LLM boxes as-is.
assessment.hitl_enabled	hitl.enabled	Moved to top-level.
assessment.default_confidence_threshold	hitl.confidence_threshold	Moved to top-level.
assessment.granular.*	Dropped	Granular assessment retired → list_batch_size (see below).
assessment.max_tokens	Dropped	v0.6 always requests the model’s maximum output.
any other assessment.* key	Dropped	The entire assessment block is retired; leftover keys are ignored on read.

When both an explicit v0.6 key and a migrated legacy value are present (a hybrid config), the explicit v0.6 key wins — so a re-seeded default never clobbers a value you set in the new shape.

The migration is automatic and idempotent

You do not need to hand-edit stored configurations. The v0.5 → v0.6 transform runs at every configuration choke point:

• when a configuration is read and validated (IDPConfig validation),
• when a stored custom delta is merged onto the current defaults,
• when a configuration is written/updated (normal update, Save As Version, Save As Default).

It short-circuits once a config is stamped config_format_version: "0.6" (unless legacy assessment.* keys are still present), so re-applying it is a no-op. In practice this means a v0.5.x stack that you update in place to v0.6 has its existing configuration migrated on first read — no manual step, and a disabled Assessment stays disabled.

Tip: Export your current configuration before upgrading (View/Edit Config → Export) as a safety backup.

Related migrations and documentation

• Granular Assessment Retirement — the companion change that replaced granular.* with list_batch_size.
• Extraction & Confidence — the consolidated extraction, confidence, and geometry guide for v0.6.
• Confidence Assessment — the three confidence modes, model, and prompts in the current shape.
• Human Review — the HITL workflow (hitl.*).
• v0.4 → v0.5 Migration — the earlier unified-pattern migration.