The problem
Naive NL2SQL dumps the whole schema into the model. On a real database that is a 413 — token limit, not a clever answer. Desktop GUIs don’t help either: they run SQL, they don’t know which twelve tables the question actually needs. Teams behind a VPN often have a Prisma file and no live connection at all.
The approach
We treated the schema as a graph, not a blob. Every dialect — Postgres, MySQL, SQLite, plus CSV via DuckDB in the browser — normalises to one contract. Before the model sees anything, a local embedder ranks tables, walks one foreign-key hop, then clamps to a token budget. Offline import (Prisma, DDL, JSON) is a first-class mode, not a demo.
The solution
A Vue 3 studio (Monaco, results grid, ERD, copilot) talking to a Fastify API. Connections are per-user. Read-only guards sit on the same path as the AI. Groq writes, explains, and fixes SQL. Embeddings stay local so we don’t pay a second vendor for vectors. Writes don’t sneak through because someone phrased the prompt aggressively.
The result
The studio runs against live databases: connect, prune, generate, explain, copy out. Large hub-and-spoke schemas no longer blow the prompt. You can import a shape with no live DB, get SQL, and take it back to the real engine. Public demo is still coming. The product already does the job.