Semantic Clustering
Shards are semantic grouping values that can link elements (memories) together.
But why?
Embeddings cant capture things like:
- time-based context
- emotional tags
- relational context (which chats are related, who said it, which persona was active)
So you need a way to ensure you are including elements (memories) that may be related but not caught by RAG.
I could just be re-inventing the wheel here. This just might be normal embeddings. I have to think more.
Also now that CAG is a thing, I may have to re-think my entire approach.