Skip to content

AI health checks

TendrilGrow can analyze a camera snapshot of your plants with a vision-capable AI model and return an agronomy-style report.

What a check produces

Each check returns:

  • A health score from 0–100 (quality-first agronomy scoring).
  • Observations about the plants and canopy.
  • Issues the model believes it sees.
  • Recommended actions.
  • A dynamic feeding schedule suggestion.

Results populate the AI health entities (score, summary, feeding schedule, last check, and a critical alert) and are stored with a retention window.

Providers

Select a provider per grow space:

Provider Notes
None AI disabled for this grow space.
Gemini Google Gemini.
OpenAI OpenAI.
Ollama Self-hosted/local models via an Ollama endpoint.

After you enter credentials (API key or endpoint), TendrilGrow discovers the available models so you can pick one. If discovery fails, you can enter a model name manually.

Requirements for vision checks

You must map a camera entity and choose a vision-capable model. Without a camera and a vision model, AI health entities are not created.

Scheduling and on-demand runs

  • Scheduled: checks run on the configured interval (default 12 hours).
  • On demand: press the Run AI Health Check button, or call tendrilgrow.run_ai_health_check.

Critical alerts

If a score is at or below the critical threshold (default 20), TendrilGrow raises a persistent notification and, if configured, calls your notify.* service. Alerts are de-duplicated so you are not spammed.

Stage-aware objectives

Scoring adapts to the grow stage set on the Growth Stage helper:

  • Mothers are judged on health and structure (never flowered).
  • Clones are judged on rooting.
  • Flowering stages are judged on quality.
  • Dry/cure stages are judged on drying/curing rather than reservoir chemistry.

Per-stage reservoir targets (pH, EC, VPD) calibrate the prompt; post-harvest stages fall back to best-practice guidance.

Grounding with cultivation context

The advisor is grounded by the editable cultivation-context helpers (strain, week in stage, targets, reservoir volume, makeup water type, nutrients, flush status, and more). Keeping these accurate improves the quality of the advice. See Entities.

History and retention

Each result is persisted with its timestamp and kept for the configured retention period (default 30 days).

Secrets safety

API keys are treated as sensitive: they are redacted in diagnostics and logs. Never paste real keys into issues or discussions.