When developers start building Retrieval-Augmented Generation (RAG) pipelines or AI agents, they often fall into a common trap: assuming a vector database is just a key-value store for floating-point numbers.
You pass text to an embedding model, store the resulting 1536-dimensional array, and run a cosine similarity search when a query comes in. In tutorials, this works flawlessly. In production, it breaks down quickly.
Real business applications rarely ask for pure semantic similarity. They do not need "anything that means something similar to this phrase." They need "support tickets similar to this issue, written in English, updated within the last 90 days, which the current user's role is allowed to view."
Getting semantic meaning, metadata rules, and access permissions to work together in a single query is where the real engineering lives.
The Problem with Pure Vector Search
A traditional relational database performs exact-match filtering ( WHERE user_id = 123 ).…