https://discord.gg/botpress logo
#👀feature-requests
IA CONVERSATION
# 👀feature-requests
e

eager-piano-43672

09/29/2023, 9:06 AM
Is there a way for AI to analyze my conversations with users on my various social media platforms like WhatsApp and adapt over time to how I typically respond to them?
c

cool-policeman-94412

09/29/2023, 10:19 AM
No. Or yes - ok. Yes, there is. No, you do not want to pay, for it, neither the development, nor the processing time to finetune the llm like every week or so with "hardly" and new data. Also you need a framework different than botpress - one that allows larger context. it IS possible, but I would not go for "over time" (maybe adjust every 3-6 months), and it it is a major project with a kb and fine tuning for the speech style, as you may run into limits what you can do there with context injection
e

eager-piano-43672

09/29/2023, 10:53 AM
Alright, I see. In the event that we still want to proceed with it, what would be the costs in terms of budget and time? And what would be the appropriate infrastructure for doing it?
c

cool-policeman-94412

09/29/2023, 11:17 AM
No idea. First, any fine tuning rules out botpress and voiceflow - botpress is hardcoded to some models, voieceflow too , none supports custom models. Second, that would require significant preparation - do you even HAVE enough data? It owuld require an investigration what you MEAN by "typically respond" - maybe it is possible, using a 16k GPT 3.5 model, to get away with a good database and a large prompt defining the style of answer you give. I have worked with that in the past - it works, until it does not anymore, then it is down to prepare a LOT of data (thousands of conversations) for the AI to naturally get the style and STILL use likely embeddings to handle the knowledge part. The first is a 4 digit budget, the later can easily reach 5 digits - the training being the smaller problem, data preparation and validation being the main one. Note that Bottpress itself is self-limiting here- the way they use knowledge base does not allow for anything more complex, I think they are hardcoded AT THE MOMENT to the small gpt3.5 model, that gives them VERY little headroom for complex prompts. Just as note - that one has 4000 tokens, I work on full Ai agents bsides, and here 4000 token is the input limit for the personality and profdession section.
e

eager-piano-43672

09/29/2023, 11:46 AM
okay, thanks!
How to integre API in the bot?
c

cool-policeman-94412

09/29/2023, 1:15 PM
Ah, depends what you mean with API -it is tricial to send a http request to an external party. More an intelligence test to find a tutorial than anything else - it is not like it is not documented. Doing anything more - like replacing the model - is simply not possible.
c

crooked-van-25152

10/02/2023, 11:38 PM
You're a rockstar @cool-policeman-94412 🙂 did you apply to become one of our Botbassadors?