At least fullHD video/imgae is possible these days, it's unclear architecture diagram compared to text on the left 4:20

@AI-Makerspace

Ай бұрын

We appreciate this feedback Aravind! Zoom outputs in 360p - a lesson we learned this time around! We'll be a bit more intentional next time!

@SDAravind

Ай бұрын

4:44 I have been working on a similar implementation but sometimes the use case is complex and the end user wants to ask anything that's possible from the knowledge base. They want to ask analytics questions and also get answers from the vector database. I have done query planning and now I'm trying to implement routing. It would be ideal to know based on the sub-question that can be answered either by SQL or Vector db alone or both. I'm using userinstuctor library for query planning. RAG doesn't always give required context as content might have duplicates. Any thoughts around this?

@nav13n

Ай бұрын

@@SDAravind In case the query is very complex, one thing that maybe of help is to break it down and rephrase it into smaller [subqueries, tools] tuples as part of the query planning step, answer individual queries with respective tools and then aggregate results with a finaliser module. However, I believe it's easier said than done in practise and might need a bit of LLM fine-tuning to achieve this. I would start with a chain of thought kind of prompting with few shot examples as a starting point for this subquery, tool/intent decomposition. Once comfortable, this same technique can be used to generate synthetic data for fine-tuning a model that can directly structure the queries into a set of subqueries fit for a particular tool use.