Order Ledger: WhatsApp order chats to a clean order book, offline

What they built

Home bakers, tiffin services, tailors and small shops take orders over WhatsApp and copy them into a notebook, and a week later nobody remembers who paid the advance. Order Ledger takes a pasted or exported chat and extracts the customer, phone, items, quantities, delivery date and time, total, amount paid and status.

The seller checks the result on a review screen and saves it. The app then shows a schedule grouped by day with today and overdue markers, a pending-payments view with a button that copies a polite WhatsApp reminder, and an all-orders table with CSV export for the accountant.

Customer names, numbers and prices are private, so everything runs on the seller's own laptop: the server listens only on 127.0.0.1 and orders live in a local JSON file. The author built it for the Hacktoberfest Open-Source AI Challenge and published it under the MIT license.

Setup

• Ollama serving qwen2.5:7b (4.7 GB) by default, llama3.2:3b for comparison
• Node.js with no packages (built-in http and fetch), one HTML file for the UI, a JSON file for storage
• A small evaluation harness (evaluate.mjs) over 18 hand-labelled synthetic chats

How it works

1. Pipeline: convertChat() turns a WhatsApp export into C:/S: lines, annotateDates() rewrites words like kal or nale into ISO dates, Ollama extracts the order at temperature 0 with JSON-schema output, and clean() validates phone, units and dates on the server.
2. Ollama's format field holds a JSON schema with enums for unit and status, so the reply always parses and uses only allowed values.
3. Relative-date arithmetic and phone-number cleanup were moved into code because the small model was unreliable at both.

Results & numbers

• qwen2.5:7b on the 6 fresh chats: 5/6 orders correct, 96% of fields, after the date fix (3/6 before)
• llama3.2:3b got only 42% of items right on the 12-chat set, so it was not shipped
• About 14 seconds per order with the 7B once loaded (about 50 s for the first order); the 3B takes about 7 s
• At temperature 0, re-running an unchanged prompt gave identical results, so every prompt change could be A/B tested

Gotchas

• Adding a fourth few-shot example dropped the 7B from 10/12 to 7/12 correct, and made the 3B invent a pineapple cake order
• A date table in the prompt listing next Tuesday made the model answer the wrong date for today and tomorrow
• The schema guarantees the shape, not correctness, which is why the review screen exists
• The author tuned on both test sets and calls the numbers indicative, not a benchmark

“The schema guarantees the shape of the answer, not that it's right.”— Khushighosh

Do it yourself

1. ollama pull qwen2.5:7b
2. Clone the repo and run node server.mjs (nothing to npm install)
3. Open http://localhost:3000 and paste a WhatsApp chat or export
4. Review the extracted order, fix any field, and save it

ollama pull qwen2.5:7b
git clone https://github.com/Khushighosh/Order-Ledger.git
cd Order-Ledger
node server.mjs      # then open http://localhost:3000