Hi everyone,
we’re building a company-chatbot, tasked with providing specific information around our product.
We started out using Davinci, but as we’re not providing thousands of pages of data and getting a lot of “halluzinations”, I was wondering, if fine tuning is the right approach.
I read a bit about embedding and many shot prompting, but am not quite sure how to coose yet.
I’m sorry for the vague question, I’m not as well versed in this field.
Fell free to ask for more information/details.
Hi @LizzzTheFox, welcome to the forum. As you are looking at a chatbot with product specific information, an embedding + few-shot would be your best bet. The embeddings would allow you to use your product specific data while the few shot will allow for the bot to answer certain questions in a specific manner.
Another recommendation I would make would be to compare a fine-tuned da-vinci model vs a gpt3.5 turbo model to see which one suits your need the most.
Tinkering with the prompt, a few shots along with a low temperature is probably the best way to try and avoid hallucinations, though the exact values/text will change from case to case, so it will be upto you to play around.
Hi @udm17 ,
thank you for your answer, it’s definitely helpful.
Do you happen to know some good ressources for learning about the embedding process and properly utilizing the many/few-shot approach?
I feel like I’d need a guiding hand and to see it done on an example, so if you know some courses/blogs or anything, I would greatly appreciate it.
The keyword you need for your own knowledgebase from which AI answers can be given is “vector database”.
The knowledge is broken into reasonable sized chunks, that then are added to a database, along with a multi-dimension AI determined vector that can be matched against user input for entailment.
OpenAI seems to have deleted their own “cookbook” document. We’ll send you to a common one, Pinecone, where you can read about what it is: https://docs.pinecone.io/docs/overview