#Make a chatGPT like bot that works like Copilot

32 messages · Page 1 of 1 (latest)

weak lagoon
#

I'm writting this thread for guidance.

My company recently adopted this (private) low-code platform and is looking to gradually shift all it's applications towards it, but before doing this the platform gotta be as flexible as their custom made applications (84 of them). It allows custom made blocks on Vue.js/Golang/... to be made and injected inside to act as UI components.

During my trial these last weeks, i made some "blocks: by myself and also MORE IMPORTANTLY used ChatGPT and Copilot to help make others. Now the important part, it took me a bit to feed all the necessary knowledge to ChatGPT for it to start suggesting what it's actually needed for.

Then went on and used enbedding on with OpenAI models, also did some finetuning, let's focus on the embedding aspect. It was succesful, i scraped the github documentation of the platform, and made index files the models go through to compare vectors.

Now the model answers well whenever asked questions about anything, it also adds predictions to have more complete answers, but my biggest issue is : It is unable to generate code when asked to, which serves the whole purpose of this project.

What am I missing ?

weak lagoon
muted agate
#

I made a plugin using a tutorial just to test code gen with our codebase and if you ask to generate code using the plugin, it does well. The tutorial was this one: https://developer.mantiumai.com/docs/openai-plugins.

Plugins are also moving to beta for Plus members if you don't already have access

weak lagoon
#

hello @muted agate it uses embeddings right ?

#

also have u tried generating code with it ?

muted agate
weak lagoon
#

dude thats my internship subject and ive been trying to reach something like this for weeks

#

no joke

#

was this made through embedding ?

#

also embedding models are different than the chatting models, so how do you get to input ur data or knowledge in with or without embedding ?

muted agate
#

Yea I used the service to import the data clean, split and embed the text. Then sent it to a Vector DB (they have a managed Redis you can use) and created my plugin. If you don't currently have Plugin access, signing up for Plus should give it to you something this week as per their announcement.

weak lagoon
weak lagoon
#

because until now for me they seem like two distinct worlds

#

and what im stuck on is merging what i embeded with the chat model

weak lagoon
muted agate
#

Just the embedding model. You can use the chat model as well to do what you need with text.

image.png image.png
weak lagoon
muted agate
#

I used Mantium to do the full end to end from getting my data ready, embed it, ship it to a Vector DB (Redis) and to create my plugin. Once it was created, I interact with it at https://chat.openai.com/

weak lagoon
#

im looking for the back end of things tho

#

how this embedding to going to the chat models goes

muted agate
#

They use ada 002 to do the embeddings.

weak lagoon
#

because the embedding model is unable to answer stuff past : Classification - simple Q & A - text retrieval...

#

if u managed to make code using the pdf u upload thats not the embedding model answering

#

thats gpt3.5 or gpt4

#

i hope u get what im trying to say ?

muted agate
#

That's what the plugin does: It takes the data you embedded from the Vector DB and surfaces the documents up to Chat GPT which uses the context to generate the response. The embedding model only embedded the data, once it's done, the query is using either GPT 3.5 or GPT 4 BUT it now has the context of your own data when you use the plugin

weak lagoon