Writing
RAG probably doesn't need a vector database
Exploring relational and lexical retrieval as an alternative starting point for many RAG systems.
A large amount of retrieval work can happen before embeddings become necessary. SQL full-text search, corpus statistics and explicit relational structure can narrow a corpus surprisingly well.
This note is a starting point for a broader argument: use deterministic structure first, then apply language models where ambiguity remains.
Why this is interesting
Vector retrieval is convenient, but convenience can hide assumptions about what information should be considered related. A more explicit retrieval pipeline can be easier to inspect, benchmark and improve.
