How to give domain knowledge to chatGPT


I am thinking about how to give my domain knowledge to text-davinci-003 and answer the questions based on that information.

I understand that it is clearly stated here that finetuning is not about giving knowledge.
However, chatGPT’s general ability to answer questions is amazing, and I would like to somehow give them domain knowledge and cover areas other than domain knowledge.

Since chatGPT also takes context into account when answering, I think if I put the knowledge to be given (about 300 words) and a question about that knowledge in the last prompt, I can get the desired result. I have tried this and this worked.

However, I am at a loss as to how to provide this to the user.
If we try to force it, we would have to send about 2MB of text at the same time each time the user wants to ask a question, and include the user’s question at the end.
However, this would take a lot of time and most importantly, TOKEN would be outrageous.

Is there any better way to do this?

1 Like

You would have to use GPT and not ChatGPT

You would need an embedded database or do fine-tuning (You cant do fine-tuning with ChatGPT)

If you did embedding, you find the context you need from your big set of text (only small part you require) and then add it as the context (as you are doing now)


In terms of specific model names, what would that be?

text-davinci-003 for the final question and running the chatbot

text-embedding-ada-002 if you went down the embedding route

davinci (without numbers) if you went the fine-tuning route

1 Like

Did you discover what you needed in the end?

I am in the same position and want to add new knowledge to a model… unsure how

1 Like