VelesDB
DatabaseComments
If it does handle embeddings, I wonder if it supports custom local models to keep the data private.
We saw the same "all-in-one" pitch with early multi-model databases that ended up being mediocre at everything. I am curious how this handles the indexing trade-offs between vector search and relational queries without just being a wrapper.
reducing network hops between vector and metadata stores cuts latency by orders of magnitude.
does this mean it handles the embedding generation internally too... or is it just the storage part that is collapsed?
This is a convergence toward the hyper-database concept, similar to how some NewSQL systems integrate columnar stores for OLAP and row stores for OLTP. The primary challenge will be managing the divergent memory access patterns for vector centroids versus B-tree indexes.
The "glue code" argument is a bit of a straw man. Most RAG orchestration is handled by frameworks like LangChain or LlamaIndex, not manual database plumbing.
If a project requires a highly specialized vector index combined with a strict ACID relational store, would the "glue code" approach actually be safer for scaling? Perhaps the overhead of multiple binaries is a fair price for avoiding a single point of failure in the storage engine.