linviaz / chatbot_demo_with_GPT3_API

Chatbot Demo with OpenAI GPT-3 Language Model API

Three components for this chatbot demo with openAI GPT-3 language model API:

• OepnAI Language Model API
• Frontend chatbot UI template
• System configuration

Step-by-step configuration

The following guideline is for linux, WSL2, or MacOS operating systems.

• clone this repository
• change directory to /config and update the system variable OPENAI_API_KEY value in the add_openai_api_key.sh file
• make sure the add_openai_api_key.sh file is executable by running the following command in terminal: chmod +x add_openai_api_key.sh
• execute the config shell script by running ./add_openai_api_key.sh
• validate the setup of creating an OPENAI_API_KEY system variable by echo $OPENAI_API_KEY

without docker

• create new conda environment with python 3.10 by running conda create -n chatbot python=3.10 in the terminal.
• in the /app directory where main.py file is, run `uvicorn app:main

Scope of Fine-Tuning and Prompt Engineering

• Fine-tuning with base models
• Embedding similarity + prompt engineering

Both fine-tuning and prompt engineering could improve the language model completion quality. Fine-tuning would update the language model parameters based on the training dataset while prompt engineering world provide additional context in the promot without modifying the pre-training language model. The key difference comparisons between fine-tuning and prompt engineering for the chatbot LLM use case are summarized in the following table:

	Fine-tuning	Prompt engineering
Training data	Normally requires large amount of task-specific training data	Small amount task-specific context data required
Accuracy	Can achieve high accuracy with sufficient training data	May require careful tuning to achieve high accuracy
Resource Usage	Requires significant computational resources and time for training	Can be computationally efficient
Potential Issues	Catastrophic forgetting, generated non-factual content	Prompt length limit and more token consumption

Preparation of training dataset

Training data embedding reference creation and provide context information.