POST /ingest -> called every 5 seconds -
client calls it with whatever is there on user's screen
POST /chat_completion - when user asks sth
type of questions(not answerable by RAG) -
- which people are in my team?
- what message i haven't yet replied to on whatsapp? give a draft reply for
those chats.
- what PRs need my reviews
- add 3 more examples
- you control the client - and can parse in whatever ways you want
- schema of ingest is for you to decide
- main objective - zoomed out view of the whole system rather than code.
- first principles, the only thing you have will be databases of your choice, an embedding model(if needed), and a LLM
- solution has to be as generic as possible
- but you can do specific things for a category of apps - e.g. conversational apps
- overall framework should still be easy to extend
- topics which might help - knowledge graph, memory - grafiti, mem0, GraphRag(you should read research paper)
- talk with AI, and maybe get more topics
- write dfs and graph stuff
Outcome -
- assuming you’re leading the effort for building above backend - a clear proposal for what you would you do and how would you do it. - written doc.
- some POC - jupyter notebook, python script - decide it for yourself. - code clarity and writing non trivial code