Parallel process GPT requests

This program can process 1000 requests to ChatGPT in 2m. It uses the gpt-3.5-turbo model by default, which you can update in the file generate_requests.py.

1000_gpt_requests_in_2m.mov

I adapted the parallel processing cookbook example written by OpenAI, which was written for the embeddings model. I adapted it to work for the Chat Completion models, which has a different request structure.

Get the program up and running

1. Clone the repo: git clone https://github.com/tiny-rawr/ZH_008_parallel_chatgpt_processing.git parallel_process_gpt_requests
2. Replace 'input.csv' with your own data. Column A should contain data for each separate request you want to make to ChatGPT. In this case, I have an input file containing details for 10 different medical specialists, one in each cell in the column. Alternatively, you can just replace the data in data.py with an array of the data you want to send in each request (not including the prompt).
3. Replace the prompt in main.py with your own prompt instructions.
4. Set your API key with export OPENAI_API_KEY="YOUR API KEY"
5. Install dependencies with pip install -r requirements.txt
6. Run the program with python3 main.py

What happens when you run the program

When you run the program, four files are generated:

• data.py turns the values in the first column of the input.csv into an array. It's saved to a file so you don't have to generate them all each time.
• example_requests_to_chat_completion.jsonl, is a list of requests that will be send to ChatGPT.
• example_requests_to_chat_completion_results.jsonl, the entire request including input and output (same as previous file but with output included).
• output.csv is a spreadsheet containing the original values passed into chatgpt in column A, and the processed by ChatGPT values in column B.