Classification state machine

Classifies uploaded PDFs using Claude on Amazon Bedrock. Reports the dominant document type + numeric confidence + page breakdown + document size as a type-agnostic DocumentProcessing-Classified event; downstream routing happens at the EventBridge bus.

Flow

States

State	Type	Purpose
FilterByCurrentYearMonth	Choice (JSONata)	Skip documents outside the current year/month path
SkipOutOfWindowDocument	Succeed	Silent no-op for out-of-window uploads
HeadObject	Task	Resolve ContentType for the format check
IsSupportedContentType	Choice	Reject non-PDF files with a typed event
ClassifyDocument	Task	Invoke classify-document Lambda
NormalizeClassification	Pass	Flatten Lambda verdict into top-level state
GetExistingTags	Task	Read current S3 object tags
BuildUpdatedTagSet	Pass (JSONata)	Merge tags, replacing any existing Classification tag
TagWithClassification	Task	Write merged tag set back to S3
EmitClassifiedEvent	Task	EventBridge PutEvents

Dynamic year/month filter

The first state uses JSONata $now(), which is evaluated fresh on every execution:

$contains($states.input.key, $substring($now(), 0, 4) & '/' & $substring($now(), 5, 2))

$now() returns an ISO 8601 timestamp (e.g. 2026-07-06T14:00:00Z), so:

• $substring($now(), 0, 4) → "2026"
• $substring($now(), 5, 2) → "07"
• Combined check → key must contain "2026/07"

Documents uploaded to an old or future year/month path hit SkipOutOfWindowDocument (a Succeed state) — no event emitted, no Lambda invoked, no Bedrock call made.

Why a Lambda for classification?

Bedrock requires PDF content as inline base64 — there is no S3 document source or URL source on the Bedrock API. A multi-page PDF base64-encodes to several MB, which exceeds Step Functions' 256 KB state-transition limit. The classify-document Lambda handles the full round-trip (S3 GetObject → base64 → Bedrock InvokeModel) and returns only the small classification verdict, keeping state transitions well within the limit.

Verdict computation

The Lambda computes the dominant type in JavaScript, not the model. The model reports per-page counts via a forced submit_classification tool call; the Lambda then:

1. Filters out UNCLASSIFIED pages
2. Sorts remaining types by page count descending
3. Picks the top type as classification
4. Computes classificationConfidence = dominantTypePages / classifiedPages

This mirrors the JSONata logic from the prior Textract-based SM so the emitted event shape and all downstream routing rules are unchanged.

Why UNCLASSIFIED is filtered

UNCLASSIFIED means "couldn't confidently classify these pages", not "a different document type". For a 47-page doc with 15 BANK_STATEMENT + 5 TRANSACTION_REPORT + 27 UNCLASSIFIED pages:

• If UNCLASSIFIED counted: dominant would be UNCLASSIFIED (27 > 15) — doc fails routing.
• With UNCLASSIFIED filtered: dominant is BANK_STATEMENT (15 of 20 classified pages → 0.75 confidence), routes correctly.

The full pageBreakdown (including UNCLASSIFIED) is still emitted for caseworker visibility.

Emitted event

Success

{
    "Source": "custom.documentProcessing",
    "DetailType": "DocumentProcessing-Classified",
    "Detail": {
        "bucket": "...",
        "key": "...",
        "object": { "size": 7390634, "versionId": "..." },
        "config": {},
        "classification": "BANK_STATEMENT",
        "classificationConfidence": 0.75,
        "totalPages": 47,
        "classifiedPages": 20,
        "dominantTypePages": 15,
        "pageBreakdown": [
            { "type": "UNCLASSIFIED", "pageCount": 27 },
            { "type": "BANK_STATEMENT", "pageCount": 15 },
            { "type": "TRANSACTION_REPORT", "pageCount": 5 }
        ],
        "documentSizeBytes": 7390634,
        "modelId": "eu.anthropic.claude-haiku-4-5-20251001-v1:0"
    }
}

Failure

• DocumentProcessing-ClassificationFailed — Lambda error, Bedrock error, S3 error, or document too large for inline classification (>14 MB).
• DocumentProcessing-ClassificationFailed — file content-type is not application/pdf.

EventBridge routing

The single Classified event is type-agnostic. EventBridge rules apply routing policy. Bank-statement extraction matches:

{
    "source": ["custom.documentProcessing"],
    "detail-type": ["DocumentProcessing-Classified"],
    "detail": {
        "classification": ["BANK_STATEMENT"],
        "classificationConfidence": [{ "numeric": [">=", 0.75] }],
        "totalPages": [{ "numeric": ["<=", 100] }],
        "documentSizeBytes": [{ "numeric": ["<=", 209715200] }]
    }
}

Adding a new document type is one new EventBridge rule + one new extraction SM — the classification SM never changes.

Timing

Step	Typical
FilterByCurrentYearMonth (JSONata Choice)	<1 s
HeadObject + IsSupportedContentType	<1 s
classify-document Lambda (S3 + Bedrock)	5–15 s
Tag operations + EmitClassifiedEvent	<1 s
Total	~5–15 s