Skip to content

VERA Overview

VERA is a verification-first OCR platform. Upload scanned documents, let PaddleOCR extract text, then review, correct, and validate the results before exporting. AI summaries via Ollama are available after validation.

How It Works

  1. Upload a scanned image or multi-page PDF.
  2. OCR runs automatically in a background Celery worker (PaddleOCR).
  3. Review the extracted tokens in the web UI. Low-confidence and flagged tokens are highlighted for attention.
  4. Correct any misread tokens and fill in structured fields.
  5. Validate the page or document. Validation is a hard gate -- nothing proceeds without explicit human approval.
  6. Summarize (optional) using a local LLM via Ollama.
  7. Export the validated text as JSON, CSV, or plain text.

Key Features

Feature Description
Multi-page PDF support PDFs are split into individual pages; each page is reviewed independently
Token-level confidence Every OCR token has a confidence score and label (trusted, medium, low)
Forced review Tokens flagged for mandatory human review before validation can proceed
AI summaries Bullet-point summaries and structured field extraction via Ollama (post-validation)
Structured fields Key-value pairs (vendor, date, total, etc.) extracted automatically and editable by reviewers
Audit trail Every action (upload, correction, validation, export) is logged with actor and timestamp
Export formats JSON, CSV, and plain text
Document type detection Automatic classification (invoice, receipt, statement) based on content keywords
Rate limiting Configurable upload rate limits via SlowAPI
Prometheus metrics Request count and latency metrics at /metrics

Architecture

VERA runs as five Docker services plus an optional backup sidecar:

┌─────────────┐     ┌─────────────┐     ┌──────────────┐
│  Frontend   │────▶│  Backend    │────▶│  PostgreSQL  │
│  Next.js    │     │  FastAPI    │     │  :5432       │
│  :3000      │     │  :4000      │     └──────────────┘
└─────────────┘     └──────┬──────┘
                           │
                    ┌──────▼──────┐     ┌──────────────┐
                    │   Redis     │◀────│   Worker     │
                    │   :6379     │     │   Celery     │
                    └─────────────┘     └──────┬───────┘
                                               │
                                        ┌──────▼───────┐
                                        │   Ollama     │
                                        │ (external)   │
                                        │  :11434      │
                                        └──────────────┘
Service Image / Framework Role
frontend Next.js Web UI for document review
backend FastAPI REST API, serves files, orchestrates tasks
worker Celery Runs PaddleOCR and AI summary tasks in the background
postgres PostgreSQL 16 Primary database (documents, tokens, corrections, audit log)
redis Redis 7 Celery broker and result backend
backup PostgreSQL 16 (sidecar) Automated daily database backups

Ollama runs externally in the shared ollama_network Docker network. VERA connects to it for AI summaries but degrades gracefully if Ollama is unavailable.

Document Lifecycle

Every document passes through a strict status progression:

uploaded → processing → ocr_done → review_in_progress → validated → summarized → exported
                │                                                        ▲
                └──► failed                                              │
                └──► canceled                              (optional, skippable)
Status Meaning
uploaded File received, queued for OCR
processing Celery worker is running PaddleOCR
ocr_done OCR complete, tokens stored, ready for review
review_in_progress Reviewer has started correcting tokens
validated All tokens reviewed, corrections applied, human-approved
summarized AI summary generated via Ollama
exported Document data exported (JSON/CSV/TXT)
failed OCR processing failed
canceled Processing canceled by user

Validation is a hard gate

Summaries and exports are only available after a document (or all its pages) has been explicitly validated. There is no way to skip the review step.

Authentication

VERA delegates authentication to Hub (the central ai.doo user management service). When Hub is configured (HUB_BASE_URL and HUB_AUTH_API_KEY), all document endpoints require a valid session. When Hub is not configured, VERA runs in open mode with no authentication.

Data Model

VERA stores five core tables in PostgreSQL:

Table Purpose
documents Top-level document metadata and status
document_pages Individual pages within multi-page documents
tokens OCR-extracted tokens with confidence, bounding box, and flags
corrections Human corrections linking original to corrected text
audit_logs Immutable event log for every document action