GPT bots in practice - the S**T that goes wrong
We’ve been building Gpt + vectorDB powered chatbots at scale since Feb. eg https://farmer.chat - a WhatsApp bot for Indian farmers, demo’d at the UN with over 40000 messages exchanged.
There’s a huge difference between a cool demo and making a valuable, retained tool.
We’ll review all the stuff that went wrong and our attempts to fix them. Eg
1. Speech reco and translation for low-resource langs
2. Dealing with vectorDB issues when:
A. You have 1000s of pdfs and YouTube videos and how to clean the data and build synthetic data
B. Supporting both vectordb and keyword search
C. Conversation summarized vectorDB queries
3. Feedback systems and how orgs iterate their knowledge base
4. How to build “Golden” answer comparative test data sets to measure incremental performance increases (or decreases) with every prompt, doc and/or model change.
This talk is part of the Seattle AI Tinkers Meetup: https://seattle.tinkerer.ai/p/ai-tinkerers-seattle-september-meetup
