Classification Use Cases (CLS)¶

Module Purpose: Automatically categorize documents by type and content category using ML models. This module contains 7 use cases.

Use Case Quick Reference¶

ID	Title	Priority
CLS-001	Extract Text Content	P1
CLS-002	Classify Document Type	P1
CLS-003	Classify Content Category	P1
CLS-004	Set Confidence Score	P1
CLS-005	Flag Low-Confidence for Review	P2
CLS-006	Manual Classification Override	P2
CLS-007	Train Classification Model	P3

UC-CLS-001: Extract Text Content¶

Overview¶

Field	Value
ID	CLS-001
Title	Extract Text Content
Actor	System
Priority	P1 (MVP Phase 2)

Description¶

Extract text content from documents for classification. Uses native text extraction for PDFs with text layers, OCR for scanned documents.

Steps¶

Determine if PDF has text layer or is scanned
For text PDFs: Extract using pdfplumber/PyMuPDF
For scanned/images: Use OCR output (from OCR pipeline)
Clean and normalize text (remove extra whitespace, fix encoding)
Store extracted text in document record

Output¶

{
  "document_id": "doc_abc",
  "text_length": 2500,
  "extraction_method": "native",
  "preview": "INVOICE\n\nInvoice Number: 12345\nDate: January 15, 2024..."
}

Acceptance Criteria¶

Text extracted from native PDFs
OCR text used for scanned documents
Text is cleaned and normalized

UC-CLS-002: Classify Document Type¶

Overview¶

Field	Value
ID	CLS-002
Title	Classify Document Type
Actor	Classification Service
Priority	P1 (MVP Phase 2)

Description¶

Predict the document type (Invoice, Contract, Report, etc.) using ML classification.

Document Types¶

Type	Description
invoice	Bills, receipts, payment requests
contract	Agreements, NDAs, legal documents
report	Analysis, summaries, reviews
letter	Correspondence, memos
form	Applications, questionnaires
id_document	Passports, licenses, IDs
receipt	Purchase receipts
other	Unclassified

Steps¶

Retrieve extracted text
Preprocess text (tokenize, truncate to 512 tokens)
Run through type classifier model
Return predicted type with confidence

Output¶

{
  "document_id": "doc_abc",
  "predicted_type": "invoice",
  "confidence": 0.94,
  "alternatives": [
    {"type": "receipt", "confidence": 0.04},
    {"type": "other", "confidence": 0.02}
  ]
}

Acceptance Criteria¶

Classification accuracy >85% on test set
Processing time <500ms per document
Top 3 predictions returned

UC-CLS-003: Classify Content Category¶

Overview¶

Field	Value
ID	CLS-003
Title	Classify Content Category
Actor	Classification Service
Priority	P1 (MVP Phase 2)

Description¶

Classify the content category (Finance, Legal, HR, etc.) independent of document type.

Categories¶

Category	Examples
finance	Invoices, budgets, financial reports
legal	Contracts, compliance documents
hr	Employee records, policies
operations	Procedures, manuals
marketing	Brochures, campaigns
technical	Specifications, documentation

Output¶

{
  "document_id": "doc_abc",
  "predicted_category": "finance",
  "confidence": 0.88
}

Acceptance Criteria¶

Multi-label classification supported
Hierarchical categories supported (future)

UC-CLS-004: Set Confidence Score¶

Overview¶

Field	Value
ID	CLS-004
Title	Set Confidence Score
Actor	System
Priority	P1 (MVP Phase 2)

Description¶

Calculate and store confidence scores for all classifications.

Confidence Levels¶

Score	Level	Action
>0.90	High	Auto-apply
0.70-0.90	Medium	Apply, optional review
0.50-0.70	Low	Apply but flag
<0.50	Very Low	Queue for manual review

Acceptance Criteria¶

Confidence calibrated (predicted 80% = actual 80%)
Thresholds are configurable

UC-CLS-005: Flag Low-Confidence for Review¶

Overview¶

Field	Value
ID	CLS-005
Title	Flag Low-Confidence for Review
Actor	System
Priority	P2

Description¶

Queue documents with low classification confidence for human review.

Steps¶

Check confidence against threshold
If below threshold, add to review queue
Notify reviewers of pending items
Track review status

Output¶

{
  "document_id": "doc_abc",
  "review_required": true,
  "reason": "confidence_below_threshold",
  "confidence": 0.45,
  "predicted_type": "contract"
}

Acceptance Criteria¶

Low-confidence items queued automatically
Review queue UI accessible
Reviewers can approve or correct

UC-CLS-006: Manual Classification Override¶

Overview¶

Field	Value
ID	CLS-006
Title	Manual Classification Override
Actor	User
Priority	P2

Description¶

Allow users to manually correct or set document classification.

Steps¶

User views document and current classification
User selects correct type/category
System updates classification
Correction logged for model retraining

Acceptance Criteria¶

Manual override possible for any document
Override history tracked
Corrections available for retraining

UC-CLS-007: Train Classification Model¶

Overview¶

Field	Value
ID	CLS-007
Title	Train Classification Model
Actor	ML Engineer, Admin
Priority	P3

Description¶

Retrain classification models using accumulated labeled data.

Steps¶

Export labeled documents
Split into train/validation sets
Fine-tune model on new data
Evaluate on holdout set
If improved, deploy new model

Acceptance Criteria¶

Automated training pipeline
Model versioning
A/B testing between models

← Back to Use Cases | Previous: Deduplication | Next: Tagging →