AI Assistant Optimization

When creating an AI assistant, there are three main factors that influence the quality of its responses:

• Descriptive words or prompts;
• Data quality of associated datasets;
• Quality of data after manual inspection and annotation.

Therefore, optimization can be divided into three parts: prompt optimization, dataset optimization, and manual inspection & annotation optimization.

Prompt Optimization

Prompts are natural language inputs that help AI understand the intent of the request.
In general, it involves clearly stating:

• Who the AI is;
• What it should do;
• How it should do it;
• To what extent it should do it;
• The goal it should achieve.

Example Reference:

# Role
GitMind Customer Support, specializing in resolving technical and daily usage issues for GitMind users.

### Skills
– Technical Troubleshooting: Proficient in diagnosing and resolving technical issues.
– User Support: Skilled in guiding users through daily usage problems effectively.
– Communication: Excellent communication skills to interact with users.

### Goals
1. Provide timely solutions to technical problems faced by GitMind users.
2. Offer guidance on daily usage issues to enhance user experience.

### Constraints
1. Maintain a professional and courteous tone while assisting users.
2. Ensure all solutions are accurate and easy for users to follow.

### Output Format
– Detailed troubleshooting steps for technical issues.
– User-friendly tips and suggestions for daily usage problems.

### Workflow
1. Receive user queries and identify the nature of the issue.
2. Provide step-by-step instructions or solutions to resolve technical problems.
3. Offer guidance on using GitMind features effectively for daily tasks.
4. Seek feedback from users to ensure problem resolution and satisfaction.
5. Document resolved issues for future reference and improvement.

## Greetings
As the GitMind Customer Support representative, I am here to assist you with any technical or daily usage issues you may encounter while using GitMind. Whether you need help troubleshooting a technical problem or seeking guidance on maximizing your daily usage, feel free to reach out. Please describe the issue you are facing, and I will provide you with detailed solutions and tips to enhance your GitMind experience. Thank you for choosing GitMind Customer Support. How can I help you today?

Additional Tips

Use the Primary Language of the Dataset in Prompts:

• If most of the dataset is in English, use English prompts to target overseas users.
• If most of the dataset is in Chinese, use Chinese prompts to target domestic users.
• If the dataset contains images, include constraints for image display in the prompts.

Use the CRISPE Prompt Framework:

• Capacity and Role: Define the AI’s role.
• Insight: Provide background information and context.
• Statement: Specify the tasks for the AI.
• Personality: Define the style of the AI’s responses.
• Experiment: Set constraints for the AI.

If it’s a customer service bot, include company or business information in the background, such as business details and contact information.

To Make AI Responses More Human-Like:

• Define the AI’s personality and response tone.
• Ensure responses match the AI’s role identity and are given in the first person.

Dataset Optimization

The optimization of the dataset involves importing more precise and relevant data. By constructing an accurate dataset, we can link it to the AI assistant during its creation. This enables us to ask dataset-related questions during chats after the AI assistant is successfully created, allowing AI to provide precise answers based on the imported data.

1. For the imported data, use complete question-answer pairs whenever possible. This means organizing typical questions into question-answer pairs beforehand, ensuring each pair is semantically coherent and accurately expressed. See the example below:



2. If the data is in table or image format, manual processing is recommended. Each data chunk should include header information or be directly converted into text paragraphs.

The chart below for example:


It can be processed into the following format:


3. The imported data will be automatically segmented. If you observe any irrelevant content in a segment, you can manually delete or edit it.


Simply modify the answer and click【Confirm】.

Manual Inspection & Annotation Optimization

Now that the prompts and dataset are confirmed, you can proceed to【Session History】to test the Q&A pairs you’ve configured. If the answers are unsatisfactory, simply click【Details】, then【Admin notes】, select the dataset you wish to modify, and rewrite the answers as needed.

With this, the manual inspection and annotation process is complete. The AI customer service bot will learn the newly annotated answer and respond accordingly the next time the same question is asked.