title	ORDO
colorFrom	red
colorTo	gray
sdk	docker
app_port	7860
pinned	false

ORDO

ORDO is a local full-stack application for digitizing handwritten, audio, and live speakerphone pharmaceutical orders into reviewed Excel workbooks.

Structure

• frontend/ - React, Vite, TypeScript, TailwindCSS, shadcn-style components
• backend/ - FastAPI, PaddleOCR integration, optional local Whisper, Gemini validation, RapidFuzz matching, openpyxl outputs

Local setup

npm install --prefix frontend
python -m venv backend\.venv
backend\.venv\Scripts\python -m pip install -r backend\requirements.txt

For faster smoke testing without installing PaddleOCR, use:

backend\.venv\Scripts\python -m pip install -r backend\requirements-smoke.txt

Audio uploads use faster-whisper from backend/requirements.txt. It runs locally. Set WHISPER_MODEL=tiny, base, small, or another faster-whisper model size in .env if you want to trade speed for accuracy.

Place the real product master workbook at backend/data/master.xlsx. The April 2026 order format uses headers on row 2; ORDO reads PRODUCT, PACK SIZE, PACK TYPE, PTS, and DIVISION, then caches extracted products in backend/data/products.json.

For text-based PDFs, ORDO parses the PDF table directly and ignores HSN, GST, MRP, rates, discounts, totals, and serial numbers when extracting quantities. For handwritten images, ORDO first tries PaddleOCR. If GEMINI_API_KEY is set, Gemini is used as a vision fallback for unsupported handwriting or image-only documents. Without either OCR dependency or Gemini credentials, unknown images return a clear processing error instead of generating a wrong filename-based row.

Phase 2 order modes:

• Handwritten Order supports one image, multiple images, PDFs, and image batches. OCR results from every file are merged into one order, duplicate products are detected, and quantities are aggregated.
• Audio Order accepts m4a, mp3, wav, and aac, then transcribes with local Whisper and filters the transcript into product order events.
• Live Voice Order captures speakerphone audio in the browser when supported, or accepts manual live transcript chunks. The live event layer supports add, increase, decrease, update, and remove actions while the review table remains editable.

The conversation filter ignores greetings, payment/logistics discussion, HSN, GST, MRP, rates, discounts, totals, and serial numbers before matching products.

Generated workbook filenames use date-month-year format: Order_DDMMYYYY_HHMMSS.xlsx and Order_Items_DDMMYYYY_HHMMSS.xlsx.

Run locally

backend\.venv\Scripts\python -m uvicorn app.main:app --app-dir backend --reload
npm run dev --prefix frontend

Frontend: http://localhost:5173

Backend health: http://127.0.0.1:8000/health

Deploy on Render

This repo includes a root Dockerfile and render.yaml for a single Render web service. The Docker build compiles the Vite frontend, copies it into the Python image, and FastAPI serves both the ORDO website and /api/* from the same onrender.com URL.

Render setup:

• Create a new Blueprint or Docker Web Service from the GitHub repo.
• Use the root render.yaml or root Dockerfile.
• Health check path: /health.
• Add GEMINI_API_KEY as a secret environment variable if Gemini fallback is needed.
• Keep WHISPER_MODEL=tiny for faster audio transcription on small instances.

Deploy on Hugging Face Spaces

Use a public Docker Space for a fully free single-service deployment. The Space metadata at the top of this README tells Hugging Face to build the root Dockerfile and expose ORDO on port 7860.

Recommended Space configuration:

• SDK: Docker
• Hardware: CPU basic
• Secrets: GEMINI_API_KEY
• Environment: WHISPER_MODEL=tiny
• Health check path: /health