Document Extraction Demo (React/Next.js + Flask)

Atlas Extract is a demo app that uploads invoices, extracts structured fields with an LLM, and stores results in a SQLite database that mirrors SalesOrderHeader and SalesOrderDetail.

Architecture

• Frontend: Next.js (React) UI with upload + editable extraction view.
• Backend: Flask API for file ingestion, LLM extraction, and SQLite persistence.
• Database: SQLite tables for SalesOrderHeader, SalesOrderDetail, plus a Documents table for invoice metadata.

Setup

Backend

cd backend
python -m venv .venv
. .venv/Scripts/activate
pip install -r requirements.txt

Create an env file:

copy .env.example .env

Update LLM values in .env:

• LLM_API_KEY=...
• Optional: LLM_BASE_URL (defaults to https://api.openai.com/v1)
• Optional: LLM_MODEL (defaults to gpt-4o-mini)
• LLM_PROVIDER=openai_compatible

For offline demos, set LLM_PROVIDER=mock.

Run the API:

python app.py

Frontend

cd frontend
npm install

Create an env file:

copy .env.example .env.local

Run the UI:

npm run dev

Open http://localhost:3000

Demo Flow

1. Start backend and frontend.
2. Upload a sample invoice from sample_invoices/ or the provided Sales Invoice.png.
3. Inspect extracted metadata + line items, edit if needed.
4. Click Save to Database and watch the Database Snapshot update.

Using data.xlsx

If you want to preload the SQLite database with SalesOrderHeader/Detail rows from data.xlsx:

cd backend
python scripts/import_excel.py --reset

Scaling Strategies (talking points)

• Throughput: Move extraction to a worker queue (Celery/RQ) with autoscaling workers; return job IDs and stream results to the UI.
• Latency: Cache repeated vendors/templates; store OCR text to avoid reprocessing.
• Document Types: Use a routing layer to detect document type and apply tailored prompts or fine-tuned models.
• Reliability: Add validation rules + confidence scores; surface low-confidence fields for human review.
• Production: Containerize services, add Postgres, use object storage for originals, and add observability (metrics + tracing).

Notes

• Documents stores invoice metadata (bill/ship to, terms, totals). SalesOrderHeader + SalesOrderDetail mirror the Excel schema.
• sample_invoices/ files are shaped to resemble SalesOrderHeader/Detail values for easy comparison.