Hi,
I have requirement to let our Chat respond with an exact answer from our knowledge base first before answering on its own. I am able achieve this using pure OpenAI completion using function calling.
What I did was define each Question as a function and OpenAI maps it whenever it thinks it needs a function to answer. So far it works.
However, it is not consistent since OpenAI decides if a function needs to be called or not. So, there are times that it answers with our pre-defined answer, sometimes it does not and uses its own knowledge. So it is not perfect. I am finding a solution to force it to use my knowledge base first before using OpenAI’s stock knowledge.
Some of my functions have external API calls to respond to questions so I am not able to try the pure embeddings approach. If embeddings + function calling can be combined, that would be ideal.
Here is a snippet of the configuration passed to the Completion API:
const initialResponse = await openai.chat.completions.create({
model: 'gpt-3.5-turbo-16k-0613',
messages,
temperature: 0,
stream: true,
functions,
function_call: 'auto'
})
I’m open to ideas which I can try out since I am new to this also.
Thanks!