Day 2 - Knowledge Graphs & Voice Agents

Yesterday I continued working on my knowledge graph spec and had a detour with a voice agent topic.

Knowledge Graphs

I think most people think of knowledge graphs as a more complicated solution to retrieval augmented generation (RAG). It's also a good tool to see relationships between data.

There are two steps to setting up your knowledge graph:

1. Identifying your nodes which will be the things you want to create relationships between. In my case its a wellness flashcard app so some of my nodes are 'User', 'UserGoal', 'Constraint', 'Flashcard', 'Storyline'
2. Creating relationships between your nodes. For example User -> SELECTS -> UserGoal; Flashcard -> IS_INCOMPATIBLE_WITH -> Constraint

I'm using a knowledge graph to power a recommender system. Defining the recommender system of flashcards is what I have been working on the past few days. I'm trying to determine how current and future flashcards are generated for users. For this I need to understand how a user acting (or not) on a flashcard impacts future cards that are recommended to them.

To better understand these actions, I will do a sample flashcard storyline for two different goals: 1. Build a Consistent Fitness Routine and 2. Stay Mentally & Socially Engaged. I think this will help me define my overall system.

Voice Agents

As for voice agents... I don't necessarily like jumping between tasks within the same day. I would rather finish something and then jump to a new domain. But sometimes we have to adjust. I had signed up for a voice agent event in Tokyo so I thought it would be worth it to switch gears a bit in the event that I wanted to ask the presenters a question.

I had previously worked on a feature that allows users to work with a voice agent to reflect on their goals. I thought it could replace the feeling of talking to a personal trainer or wellness coach. But I have doubts that people would want to have a long conversation with a clanker. So I think I will keep the feature so that users can have a short daily or weekly 'Wind-Down' to learn about their data and reflect amongst themselves.

As for the Voice Agent event, it was 95% noise - but the other 5% was worth hearing. The last speaker presented a high level overview of a product they are building (see pic above). They shared their use case for LiveKit. I had previously thought to use Vapi but I think I will use LiveKit instead when I get to building out this feature.