Practical tools for system, network, security, and operations administrators.

DBA-C0172 · AI, Vector & Analytics Databases

Text-to-SQL with LLMs

Read a prompt that asks for one SELECT. The page does not call a model and does not run the statement.

Updated 25 October 2026 · How-To

This is an original planning guide. It is not a vendor manual, an official standard, or legal advice. Example addresses use documentation ranges. Review every command before you paste it.

What this hour is for

The builder writes a prompt that names the schema you paste, asks for one SELECT, and says not to invent tables and not to emit UPDATE, DELETE, or DDL. The note says to review the statement before it runs anywhere. The page does not send the prompt to a model.

Do not paste a secret into the schema. Do not run a statement from the page. If a later tool writes a statement, a person still reads it. Untrusted text does not get concatenated into SQL.

192.0.2.10, 198.51.100.0/24, 203.0.113.0/24, and 2001:db8::/32 are documentation. example.com is a documentation name. A sample password is replaced. It is not a password you keep.

The change you can finish

Apply nothing while you are writing the note. When the lab window opens, it is one change on the database named in the ticket. Read the draft, then the database, then decide.

Rollback is the backup, the file, or the snapshot you saved first. You do not invent a rollback after the only copy is gone.

  1. Paste a schema you may share.
  2. Read the prompt.
  3. Confirm it asks for one SELECT.
  4. Do not run a statement.
  5. Stop.

Mistakes that outlive the statement

Text-to-SQL with LLMs goes wrong in a small set of ways. The list is here so the ticket can forbid them before the window.

Decisions to write down

One decision for Text-to-SQL with LLMs: The schema has no secret. The prompt asks for one SELECT. Nothing was run. And no model was called from the page.

Checks before you call it done

Confirm this before you call the change done: The schema is written. The prompt was read. Nothing was run. No secret is in the note. And no model was called from the page. If one of those fails, stop.

# One SELECT. Do not invent tables.
# Do not emit UPDATE, DELETE, or DDL.
# The page does not call a model.
# Review any statement before it runs.

Before you leave the database

Text-to-SQL with LLMs ends when the database in the prompt is the database in the ticket. If it is not, you are on the wrong database. Stop.

The draft stays a draft until a person has read the names in it. A generator does not log into the database. You delete a line that names a host, a database, or an account you did not write down.

192.0.2.10, 198.51.100.0/24, 203.0.113.0/24, and 2001:db8::/32 are documentation. example.com is a documentation name. They belong in the note. They are not values you leave on a database because a draft included them.

prompt-for-this is a reminder that a secret is missing. It is not a password, a key, or a token you save in the ticket or in git. A sample password in a draft is replaced. It is not a password you keep.

One change is the window. If two things move and the result is wrong, you will not know which one did it.

If the same step fails twice, stop for the day. Restore the copy you took first.

A second person reads the database name and the change aloud. If they cannot find them in the ticket, the note is not done.

This page is a planning guide. It is not a vendor manual, not an official standard, and not permission to touch a database you do not administer. You do not disable authentication, a firewall, or TLS to make a check pass.

Tools for this guide

These tools run in the browser on this site. They do not log into your machines.

More in this cluster

The pillar for this page, and the next published pages filed with it.