how to train gpt chat api to make custom replies in my message flow tool

SteveeJobs

Regular Member
Joined
May 1, 2023
Messages
239
Reaction score
48
I want to create an automation model with message flow, and integrate the chat gpt API with the message flow tool API, but not to get too vague, because the chat gpt Q&A may have nothing to do with my niche or services, I want to find a way to train the Artificial Intelligence to do exactly what I want, like answering and selling, also providing support.
 
I want to create an automation model with message flow, and integrate the chat gpt API with the message flow tool API, but not to get too vague, because the chat gpt Q&A may have nothing to do with my niche or services, I want to find a way to train the Artificial Intelligence to do exactly what I want, like answering and selling, also providing support.
chatGPT doesn't come with fine-tuning APIs. Most you can do is prompt engineer and one shot training in the context window.

There's an option of doing something programmatic first (like, doing some NLP to figure out that some message is a support question, related to X topic) and then feed chatGPT some context to generate some text.
These days using vector embeddings and vector search for these steps is getting popular.

If you want more fine-tuning control. Use some GPT3 model. Or host an open-source model like LLAMA2.
 
chatGPT doesn't come with fine-tuning APIs. Most you can do is prompt engineer and one shot training in the context window.

There's an option of doing something programmatic first (like, doing some NLP to figure out that some message is a support question, related to X topic) and then feed chatGPT some context to generate some text.
These days using vector embeddings and vector search for these steps is getting popular.

If you want more fine-tuning control. Use some GPT3 model. Or host an open-source model like LLAMA2.
so it is possible, for example: the customer comes to me, the flow talks about the product, makes the offer and the api could break the objections, regardless of the question, but always related to the product. It is possible?
 
omes to me, the flow talks about the product, makes the offer and the api could break the objections, regardless of the question, but always related to the product. It is possible?
I am not sure what you mean by "break the objections", can you be more descriptive?
It sounds possible. Would need a bit of clever prompt engineering ig
 
I am not sure what you mean by "break the objections", can you be more descriptive?
It sounds possible. Would need a bit of clever prompt engineering ig
"Breaking objections" would be for example:

Customer: "But how does this help my business?"

ChatGPT responds by always directing my company and my services, this was a very basic example, but you can see that I want a complete service, from the flow of messages ready, chatgpt would come in to "break objections" or provide support ( based on training).
 
You'll be able to fine-tune 3.5 turbo and 4 in the future, currently davinci is your best option: https://platform.openai.com/docs/guides/fine-tuning
ChatGPT is essentially a fine-tuned version of 3.5, there's no option to refine it further except for working on what prompts you use with it.
Or as @Ashk881 says above, you can look at open-source models - Falcon is looking really good at the moment https://huggingface.co/tiiuae/falcon-40b
Perfect friend, I'll study and talk to the programmers here at the company, thank you!
 
"Breaking objections" would be for example:

Customer: "But how does this help my business?"

ChatGPT responds by always directing my company and my services, this was a very basic example, but you can see that I want a complete service, from the flow of messages ready, chatgpt would come in to "break objections" or provide support ( based on training).
Oh yes, chatGPT can do this to some degree. you just need to prompt engineer it with some objections you can imagine and the possible counters.
The prompt window is large enough to accomodate basic examples.

If you want something more robust, you'll need to finetune a custom model with a database of your usecases and knowledgebase.
 
For fine-tuning you either do it with lalma model or old Gpt model davinchi. but based on your need, it should be enough.
 
Chatgpt won't remember after the session but other models in AI, u can train them
 
The best way is that you use the same chat of the whole message so chat gpt keep getting improved in every message
 
Back
Top