LLM assist

Adminium’s schema classification is heuristic and deterministic. It reads your types, keys, and names, and it decides. No model is involved, and none is required. Point Adminium at a database with no AI configured at all and you get a complete app.

LLM assist is a layer on top of that: it can propose better labels, spot enums your schema did not declare, suggest relations, and write descriptions. Every suggestion arrives as a reviewable diff against the heuristic baseline, never as a silent overwrite.

Two paths, one contract

	Direct API	BYO round-trip
How	Adminium calls the provider with your key	You copy a prompt, paste the response back
Network	Adminium → provider	None. Zero.
Providers	Anthropic, OpenAI, Ollama	Anything you can paste into
Metered on Adminium Cloud	Yes, with Adminium-managed keys	Never
Available offline	No	Yes

Both paths produce the same run, validated by the same schema, reviewed through the same diff, applied by the same transaction. The BYO path is not a degraded mode — it is the same contract with you as the transport.

• BYO round-trip — copy the prompt, paste the response.

What is sent

Whichever path you take, the prompt carries schema metadata and aggregates only:

• Table and column names, types, nullability
• Keys, constraints, indexes
• Counts and cardinalities

Not your rows. The single exception is opt-in and explicit: --sampling includes a capped number of example values per column, because sometimes the only way to know status holds 'A' | 'B' | 'C' is to look. It is off by default and you have to ask for it by name.

What can be suggested

Section	What it improves
labels	Human names for tables and columns — cust_ref_no → “Customer reference”
enums	Values a varchar actually holds, so it becomes a select
relations	Relationships your schema implies but does not declare
descriptions	Field help text
pages	Layout and grouping proposals

Request a subset with --sections labels,enums.

Precedence, and why nothing is ever clobbered

user > llm > heuristic

Your own edits always win. An LLM suggestion overrides the heuristic default, never a value you set yourself. Applying the same run twice writes no duplicates. Every applied suggestion records its provenance, so you can always answer “why does this field say that?” — and undo it.

Confidence

Every suggestion carries a confidence score. The review screen defaults to accepting everything at ≥ 0.8; the CLI’s --yes-above is the same control with the same default.

Below the bar is not “wrong”, it is “look at this one”. Lower the threshold and review more.

None of this is required

There is no degraded experience without it. The heuristic classifier is the product; LLM assist makes a good app slightly better at naming things. If your policy forbids sending schema anywhere, use the BYO path — or nothing at all.