florent-escribe / idsctppromptengineering

A training exercise on prompt engineering for the IDSC class of Mines Paris

idscTpPromptEngineering

A training exercise on prompt engineering for the IDSC class of Mines Paris
Jan 2024

Context :

You are trying to automate the invoice processing service of a company.
In data/pdf_documents, there is a dataset of invoices and purchase orders.
In data/ocr_results, there is the output of Amazon Textract, an OCR system, on those documents.
Your goal is to determine for each pair of Invoice/PO whether :

• The PO number is the same
• The receiver name is the same
• The vendor name is the same
• The invoice total is lower than the PO number

Instructions :

1. MODOP: Get your OpenAI API key

   • Go to https://openai.com/product
   • Create an account or login with google
   • Fill in personal information
   • Label data Prove you are human
   • Go to API keys on the left panel
   • Verify your phone number
   • A window automatically opens to create an API key
   • Give it a name
   • Save it to a local text file
   • Go to Billing on the left panel
     • Click Add payment details
     • Select Individual account
     • Fill in your payment details
     • Ask for the minimum 5$
     • Pay (6$ total with tax)
   • All set !

2. Create a .env file at the root of the project

3. Add

OPENAI_API_KEY="[your api key]"
PYTHONPATH="[the path to your projet]"

4. In the terminal, run source .env. You can check that it has worked by looking at your environment variables.
5. Create a python virtual environment in the project
6. Install project requirements by running pip install -r requirements.txt
7. In notebooks/ open prompt_examples.ipynb and run your first GPT prompts through OpenAI's API.
8. In notebooks/ open extraction_example.ipynb, extract the PO#, receiver/vendor names and total for one document.
9. In notebooks/extraction_on_dataset.ipynb, you will find code that extracts those fields on the whole dataset.
10. In notebooks/ open matching_layout.ipynb, determine for each order if the PO fields match those of the invoice.
11. In notebooks/matching_on_dataset.ipynb, you will find code that matches POs and invoices.
12. Iterate on the prompts to get better results.