How To Develop And Train AI Bot

How to build an AI Bot on GPTBots to quickly apply it in the current business.

On GPTBots, after completing the registration and authenticated as developer, you can enter the developer portal. In the developer portal, you can create AI Bots with more, more powerful and more flexible functions.

Write an appropriate identity prompt for the Bot

Role, Skills, Personality

Role: Please assign the LLM an identity role in which the LLM will think about the problem (with valid best practice).

Skills: Tell the LLM what to do and what tools or methods to use.

Personality: The tone, personality, communication style, etc. when asking the LLM to output something.

Goal

Tell the LLM of a goal that the LLM wants the Bot to accomplish. Goals should be very clear and specific.

Chain of Thought

Tell LLM what steps must be taken to think about the problem in order to achieve the goal.

MUST follow these steps to answer the customer queries
Step1 - Step2 - Step3 - Step4 - ...

Output

Clearly inform LLM of the output format and requirements you need, such as: JSON, MD format, table, emoji and other formats. Negative words to avoid, length of reply, style of language, etc.

Example

LLM is very good at imitating with examples. After providing typical examples to LLM, LLM will usually respond and output strictly according to the examples when encountering similar situations.

Prompt Example

###

Use the language that the user previously used or the language requested by the user.

You are the customer service of an e-commerce website (www.xyz.com), and you need to help customers solve problems professionally and politely.

###

MUST follow these steps to answer the customer queries.

Step 1: First determine whether the user is asking about a product, a store, or XYZ, and politely decline to respond if it is an irrelevant question.

Step 2: Please determine whether the user's question contains specific product or store information.

Step 3: If you include product and store information, list any assumptions the user made in their information, for example: Laptop X is more expensive than Laptop Y, or Laptop Z has a 2-year warranty.

Step 4: If the user has made any assumptions, please use the product information and store information to determine whether the user's assumptions are correct.

Step 5: First, politely correct the customer's incorrect assumption if the user's assumption is incorrect. Only mention or quote correct product information and respond to customers in a friendly manner.

You don't need to reply to the content of Step 1 ~ Step 4, just output the final result directly.

Upload appropriate knowledge to the Bot

GPTBots currently supports uploading knowledge to Bots in various ways (file uploading, website crawling, Q&A, text editing, etc.), and developers can easily complete the uploading of relevant knowledge and vector indexing, so that Bots can perform tasks based on the uploaded knowledge. response. But developers still need to do more work to make bots more responsive, GPTBots suggests:

• Prepare relevant materials according to the Bot's position, avoid uploading or grabbing all data without distinction;

• Prioritize the use of text and Q&A to upload data to prevent loss in the process of data recognition;

• It is strongly recommended to spend more time on designing and improving Q&A, high-quality Q&A can also be used for model fine-tuning;

• Through the "BOT Settings - Knowledge Learning - Vector Search" function, test the search results and scoring and optimize.

Train and refine Bot's responses

GPTBots provides developers with a chat record management function. Developers can view AI responses in "Bot Settings - Chat Records". For responses with poor quality, developers can provide the best response answer through the "Modify" function. When the Bot faces similar problems in the future, it can give priority to responding according to the corrected answer.

Solve complex problems by Plugin or Flow

If developers are faced with complex and special problems and conventional Bots cannot respond effectively, GPTBots recommends that developers try to use Plugin and Flow to solve the problem. Plugin can easily connect LLM with the developer's business and data to solve certain single and clear problems. Flow allows ordinary users to quickly build a workflow to solve some complex problems.

Apply Bot to business through API or external embedding

After bot debugging is completed, developers can integrate bot capabilities into their own business through widget embedding, API, etc., so as to realize the seamless AI capability experience of end users and serve the growth and development of business.