How to train the OpenAI language model in CRM?

Our development team is currently (1.06.2023) at the halfway point of implementing the language model from OpenAI in SalesWizard CRM. We plan to launch a test version of our integration soon. In this article you will learn a little more about how you can use OpenAI in your SalesWizard system.

How to train the ChatGPT language model?

In order to teach ChatGPT to answer customers’ questions and move along a certain pattern of provided knowledge, you can use several strategies. Here are some steps that can help you:

• Provide training data: Prepare a training data set containing customer questions and answers according to the expected response pattern. The more diverse examples you provide, the better ChatGPT will be able to learn.
• Using appropriate tags: You can introduce special tags in the text to help ChatGPT understand the structure of the response scheme. For example, you can use the “Question:” tag before the customer’s question and the “Answer:” tag before the expected answer.
• Identify acceptable question types: You can teach ChatGPT to recognize different types of questions, for example, questions about prices, terms, product features, etc. Identifying these types can help ChatGPT tailor responses to a specific category.
• Generate responses within a scheme: When generating a response, you can use patterns or templates that will implement the structure of the response according to the expected scheme. For example, if the response scheme includes “Introduction,” “Benefits,” and “Conclusion” sections, ChatGPT can enter these sections in the correct order.
• Testing and correcting errors: After teaching ChatGPT and generating responses, it is important to regularly test and validate responses. If ChatGPT makes errors or doesn’t deliver the expected responses, you can adjust and correct the training data, adjust the tags or correct the response patterns.
• Use of human verification: You can also use human verification to check and evaluate the responses generated by ChatGPT. Ask test questions and compare the answers generated by ChatGPT with the expected answers. In case of errors, correct the training data and continue the iterative correction process.

It is important to understand that ChatGPT is based on statistical analysis of a large amount of training data and may generate responses that are not always correct or expected. This requires an iterative approach and continuous optimization to achieve the best possible results.

Prepare questions and answers about your services, sales process, complaint process, policies, offerings. Complete the following sections.

What might an example data structure for the AI ChatGPT language model look like?

Here is a sample training data structure that you can use to teach ChatGPT to answer customer questions with a specific response pattern:

Example of a product question:

Question: what are the features of product X?

Answer: Product X offers the following features: [wypisz funkcje]

Example of a price question:

Question: how much does product Y cost?

Answer: The price of product Y is [cena] PLN.

An example of an accessibility question:

Question: is product Z in stock?

Answer: Yes, product Z is in stock.

Example of a question regarding the delivery date:

Question: when will I receive the W product I ordered?

Answer: The expected delivery date of product W is [termin].

Example of a complaint question:

Question: how can I file a claim for a V product?

Answer: To make a complaint about product V, please contact our customer service department at [numer] or send an email to [adres].

Example of a question regarding the return of goods:

Question: how do I return a purchased U product?

Answer: To return a U product, please contact our customer service department within 14 days of receiving your order. The return shipment should be sent to [adres] with information about the reason for return.

These examples show a structure in which the customer’s question is preceded by the “Question:” tag, and the expected answer is preceded by the “Answer:” tag. You can customize this structure and add more examples that are related to your specific case and expected response pattern.

Remember that the more diverse examples you provide in the right structure, the better ChatGPT will be able to learn and generate responses as expected.