Keep real data private

Problem: you must generate realistic data without exposing production values to any third party.

Solution:

# Default: nothing leaves the machine (embedded model only)
dbsprout generate --engine spec --privacy local

# Structure only may reach a cloud LLM; never row values
dbsprout generate --engine spec --privacy redacted

Keep the tier in dbsprout.toml:

[privacy]
tier = "local"

Why it works: the privacy gradient is explicit — local keeps everything on-device, redacted permits only schema structure to a cloud LLM, cloud permits sample data and requires an API key. The default heuristic engine makes no network calls at all. For training data, PII redaction is on by default (see Fine-tune a local model).