Challenge 4- 3: Automated Prompt Engineering Challenge

• 

  Rule-based Chatbot

• 

  Proposed Solution

• 

  Deployment Diagram

• 

  Genetic Algorithm

Inspiration

Traditional rule-based chatbots rely on predefined rules and patterns to generate responses. These bots follow a set of programmed instructions and keyword-based algorithms to interpret and respond to user inputs. The traditional chatbot has the following limitations:

1. Lack of context understanding
2. Limited problem-solving abilities
3. Lack of personalization
4. Incapacity to learn

What it does

Our proposed chatbot solution utilizes a genetic algorithm to generate optimized prompts based on user input and feed the best prompt into our large language model. The main highlights of this solution are:

1. Genetic Algorithm is used to generate the best prompt which makes the client’s questions easier for LLM to understand.
2. The LLM-powered chatbot will respond to clients’ inquiries in a human-like manner and generate contextual responses.

How we built it

We have developed a GA-based system for our prompt optimization to pass a filtered and optimized prompt to the LLM.

Challenges we ran into

Designing an appropriate Genetic Algorithm based on an E-commerce customer service user case.

Accomplishments that we're proud of

1. The genetic algorithm can generate better prompts given user input.
2. We built a customer service chatbot to implement our solution.
3. We built a function called "prompt guard" to filter sensitive prompts to give users a better experience.

What we learned

We learned everything from understanding the problem, designing a hypothetical solution, then implementing the solution, and pruning until the final version we are presenting.

Future work

1. Build diverse prompt datasets and introduce online learning which can help continuously update the dataset.
2. Improve fallback mechanisms and continuous monitoring.
3. Finetune the Large Language Model (LLM) according to specific user cases.
4. Enhance the chatbot interface by incorporating user-friendly functions.
5. Enhance the prompt guard's filtering capabilities using GA capabilities.

Built With

• docker
• llama
• python
• streamlit