I'm building a CRDT replicated EAV graph store with reactive datalog optimized with a cascades query planner, and uses random walks during planning of the data on disk (assume NVMEs) for cardinality estimation. The idea is to provide a structured knowledge base that can quickly run queries on relatively small data (10m-100m facts) to an LLM, and teach the model to always do research within the knowledge base using powerful analytical queries wrapped in a pure functional language for formatting results (starlark in my case). Then, you would create adapters to sync data into the system, and the data instantly becomes discoverable by the LLM to help answer questions in a structured way. The LLM could also subscribe to queries and receive delta propogations as data changes to agentically react to updates.

The knowledge base is eventually consistent across computers so you could run ingestion into a knowledge base on your desktop, and just connect and replicate it to your laptop when it's online. I've been having a lot of fun learning about Cascades query planning, and it's given me an excuse to spelunk in the Data Fusion codebase (we're both writing query planners in Rust). It turns out Microsoft put together an incredible reference about this in 2024 [1], and one of the primary authors did an awesome talk about it on youtube [2], for those interested.

And yeah, it's all vibe-coded. I've gone through 8 PoCs for cardinality estimation and 3 query planners before arriving at my current design. The ability to try stuff, rip it out, try other things; has been incredible. I would have either given up or been stuck with my first or second attempt right now without an LLM to help me quickly prototype concepts

[1]: https://www.microsoft.com/en-us/research/wp-content/uploads/... [2]: https://www.youtube.com/watch?v=GQsoNdvFbuo