Skip to content

DEVI learning

Open DEVI anywhere

Press Ctrl + Option/Alt + D in the workspace to open or close DEVI. The shortcut works from any page and keeps the current page as context. Escape closes the panel.

DEVI improves from explicit feedback, not silent prompt collection.

What is stored

When you react to an answer — hold it and choose an emoji — CloudSlash stores:

  • your question,
  • DEVI's answer,
  • the actual reaction,
  • a derived rating (up, down or neutral),
  • the provider and model name.

Credentials, API keys, emails, connection strings and resource account IDs are redacted before the interaction is stored in a dedicated devi_feedback table, outside the audit trail, with a 90-day retention window. Feedback is stored only when you rate an answer.

Export a training set

DATABASE_URL=postgres://... \
  python3 scripts/devi/export_feedback.py --org org_xxx > devi_feedback.jsonl

The export redacts again and writes prompt / completion rows with the reaction and derived rating. Merge only the explicit signals into the curated dataset:

python3 scripts/devi/export_feedback.py --org org_xxx > devi_feedback.jsonl
python3 scripts/devi/build_devi_dataset.py --feedback devi_feedback.jsonl

The builder keeps the curated golden conversations and only adds feedback rows rated up or down; neutral reactions stay as review notes, not training labels.

The honest privacy boundary

A cloud model cannot answer without seeing the prompt. When you choose Gemini, OpenAI, Anthropic, OpenRouter or Adaption, the redacted prompt crosses that provider's boundary. CloudSlash minimises it:

  • evidence is a projection (no raw resource properties, no credentials),
  • CRNs are stripped of account and resource IDs,
  • secrets and direct identifiers are replaced with [redacted],
  • the full organization context is never stored in a third-party dataset.

If the requirement is "nothing leaves my machine", use a local model endpoint. No hosted provider can truthfully promise zero data transfer.

What DEVI does not do

  • It does not execute scans, approve plans or mutate infrastructure.
  • It does not learn from prompts you did not rate.
  • It does not send credentials in chat; if you paste one, it is redacted before the provider call and you are told it was removed.