I've promised to tell you about the project that my team and I worked on during the lablab.ai hackathon.
In short, it's about fine-tuning based on external data (we used Notion as a basis).
The overall scenario is as follows:
1. Connect your knowledge base (Notion, Confluence, etc.)
2. We use embeddings to find vectors for the information and store them.
3. The user asks a question (we made it as an integration in Discord and Telegram, but it can also be a search, a field on a website, a voice assistant, or something else).
4. Profit
Learnings:
1. The hackathon is a great opportunity to try out tools that you would normally spend months learning in just a week or two.
2. The devil is in the details; many ideas seem obvious, but once you start implementing them, you find that 80% of the time is spent on various nuances, wrappers, etc. I think this is the main difference between realized products and businesses and mere ideas.
3. In the end, “adding knowledge” to the model turned out to be better through embeddings (vectors) rather than through fine-tuning.
a) it's cheaper
b) it's technically more correct. In short, each user query is processed through an embedding, a vector is found for it, and then you search your database for the closest question-answer pair to what the user asked, and you use all of this as part of the prompt to form the final answer.
Project card + video: https://lablab.ai/event/openai-hackathon/data-dreamers/i-am-ai-personalized-chatgpt
Discord for text: https://discord.gg/qy2MgAXE
Thoughts for the future:
1. Experiment with whether embeddings can understand and classify actionable suggestions like “send this link to Vasya,” understanding that action=send email.
2. One of the sources of knowledge, besides Notion, for us is chat history. I want to experiment with extracting the chat history, creating vectors for each message, and automatically classifying them. For example, identifying “the top 5 topics discussed this week,” even if they were phrased differently.
P.S. Message me if you are thinking about applying AI in your project but don't know how to approach it yet.
