Personally, I have found the best way to integrate ChatGPT and Confluence into a Chatbot is using Landbot using Webhooks to Post using REST to Openai and using Get to retrieve titles and pages from Confluence.
My game in town is to use ChatGPT to do the smart Keyword extraction from the user's input and semantics using some prompt engineering, I interpret and extract keywords from user input like "How the hell do i change my password" where it ignores idiomatic expressions and Sends a GET request to Confluence with the keywords each delimited by a + character and by adding an asterisk to the end of each keyword so it can find titles. Using the asterisk solves word matching like the word 'forgot' to find an article title with the word 'forgotten' in it. This works well for my use case and BETA chatbot for support using our Confluence KB. The final ChatGPT Post does all the summarising of the page in Confluence so it outputs it in the most readable and digestible way to read in our Chatbot, which makes it far easier to read than the whole original article in Confluence. To do that I had to strip out the HTML first for it to parse and re-write it how i wanted.
First, you need to extract the keywords from the user input, so I use Openai to do this via a POST request, and then I send a GET request with the querystring title~{keyword1*+keyword2*} so that you get back articles using the Confluence REST API.
I next output the matching results into Buttons labelled with the article titles, so the user can pick one specific relevant page article,
then I send a GET request with the pageID using another request to Confluence again,
I pass that result using some Prompt engineering as a POST again to Openai which I prompt it to then digest and output the Confluence Page content in a format that I am happy to present it in, back to the user (summarised and listing steps to take where applicable), plus I place a button to show the full article to the user via a Modal window, as the content in the REST JSON is all HTML, so they can also see the whole page article if they want to and its all well formatted. Best of both worlds then.
So far the Landbot community reckons this is a pretty smart solution. It uses ChatGPT 3.5 turbo, and will use 4, which means there is no need for training (fine tuning and embedding), plus also it requires relatively few tokens by virtue of my using ChatGPT for what its good at - language and summarising the text in the Confluence KB article, which is exactly what I wanted and required for my end-user support use case.
You must be a registered user to add a comment. If you've already registered, sign in. Otherwise, register and sign in.