adrienne
@adrienne
I think I need to give the bot the ability to choose different paths depending on the nature of the question. Path 1: semantic search over all transcripts to find the episode we talked about a certain topic, then enrich the transcript with metadata (date, YouTube url, episode title etc) For questions like: “What are some best practices for new users” “What is clanker” “What are warps” Path 2: First search across the meta data, then enrich it For questions like: “What guests were on episode 100” “How many times has DWR been on the show “ “When was the first episode” I guess the ultimate distinction is whether the primary answer to the users question is found in metadata or transcripts
1 reply
0 recast
8 reactions
Zach
@zd
You can probably get both of those to work, but I'd guess that Path 1 would solve 90% of the questions people are asking your bot. (Plz correct me if I’m wrong!) Also, if you haven't already, this post from Anthropic contains a lot of great context on the different configurations of stitching together LLMs to create agents. https://www.anthropic.com/research/building-effective-agents
1 reply
0 recast
2 reactions
adrienne
@adrienne
Thanks! I asked people to tag gmfc101 with questions today so I could see the types of questions and I got a lot of path 2 questions. I’m going to send all the questions people asked it and the replies to chatgpt and ask it to suggest ways to improve bot response. Will be traveling this week so might take a minute before I get back to it but I think I have a high level direction. The post you shared looks helpful- thanks
1 reply
0 recast
2 reactions
Zach
@zd
Oh wow! Sounds like a good learning and plan :)
0 reply
0 recast
1 reaction