Document Extraction Domain Model

Core Definition: Parsing financial statements using vision models with confidence scoring.

Data Flow

flowchart TB
    A[Upload PDF/Image/CSV] --> S[Store to Object Storage]
    S --> P[Create PARSING Statement]
    P --> B{File Type}
    B -->|PDF/Image| C["OpenRouter Vision Model"]
    B -->|CSV| D[Structured Parser]
    C --> E[Extract JSON]
    D --> E
    E --> F{Confidence Score}
    F -->|≥85| G[Auto-Accept]
    F -->|60-84| H[Review Queue]
    F -->|<60| I[Manual Entry]
    G --> J[(PostgreSQL)]
    H --> J

Confidence Scoring

Thresholds:

• •≥85: Auto-accept
• •60-84: Review queue
• •<60: Manual entry required

Supported Institutions

Data Integrity

To prevent floating-point errors:

1. •AI Output: LLM prompt requests monetary values as numbers or strings
2. •Pydantic Validation: NEVER use float for amount fields. MUST use Decimal
3. •Database Storage: Stored as DECIMAL(18,2)

Parsing Resilience

• •Bucket auto-create: storage ensures the bucket exists before upload
• •Orphan cleanup: if DB persistence fails after upload, the uploaded object is deleted
• •Stuck job supervisor: statements stuck in parsing longer than 30 minutes are marked rejected

Source Files

• •Models: apps/backend/src/models/statement.py
• •Schemas: apps/backend/src/schemas/extraction.py
• •Logic: apps/backend/src/services/extraction.py
• •Validation: apps/backend/src/services/validation.py
• •Storage: apps/backend/src/services/storage.py