Ground generation prompts with bounded real source samples #10

New Issue

2026-08-28T13:34:23Z

mptyl commented

2026-08-28 13:34:23 +00:00

Parent

What to build

Improve selected and database-wide description generation with bounded real source context while making the disclosure visible and preventing sampled values from becoming persistent operational data.

Acceptance criteria

A model request may include no more than five real rows and no more than five representative distinct non-null example values for relevant columns.
Sampling uses the binding's existing read-only data-access capability when available and never mutates the source database.
Unavailable optional samples do not fabricate values; generation can continue from catalog metadata with a safe diagnostic.
Schema names, comments, descriptions, and values are bounded and treated as untrusted prompt content separated from instructions.
Sample rows and values are never stored in runs, events, catalog metadata, API responses, or application logs.
The UI and operator documentation disclose that real source values may be sent to the selected provider before generation is used.
Tests prove the five-row and five-example limits, metadata-only fallback, read-only behavior, prompt boundary, and absence of sampled values from persistence and logs.
The future Sensitive Data Policy remains explicitly documented as required follow-up work; this ticket does not invent a partial classifier.

Blocked by

#8 — Generate selected tables and columns in structured batches

## Parent https://git.tylconsulting.it/mptyl/ThothII/issues/4 ## What to build Improve selected and database-wide description generation with bounded real source context while making the disclosure visible and preventing sampled values from becoming persistent operational data. ## Acceptance criteria - [ ] A model request may include no more than five real rows and no more than five representative distinct non-null example values for relevant columns. - [ ] Sampling uses the binding's existing read-only data-access capability when available and never mutates the source database. - [ ] Unavailable optional samples do not fabricate values; generation can continue from catalog metadata with a safe diagnostic. - [ ] Schema names, comments, descriptions, and values are bounded and treated as untrusted prompt content separated from instructions. - [ ] Sample rows and values are never stored in runs, events, catalog metadata, API responses, or application logs. - [ ] The UI and operator documentation disclose that real source values may be sent to the selected provider before generation is used. - [ ] Tests prove the five-row and five-example limits, metadata-only fallback, read-only behavior, prompt boundary, and absence of sampled values from persistence and logs. - [ ] The future Sensitive Data Policy remains explicitly documented as required follow-up work; this ticket does not invent a partial classifier. ## Blocked by - #8 — Generate selected tables and columns in structured batches

mptyl added the enhancement ready-for-agent labels 2026-08-28 13:34:23 +00:00

mptyl added a new dependency 2026-08-28 13:35:03 +00:00

#8 Generate selected tables and columns in structured batches

mptyl closed this issue

2026-08-28 16:09:42 +00:00

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Depends on

#8 Generate selected tables and columns in structured batches

mptyl/ThothII

Reference: mptyl/ThothII#10