# Portainer v3 Templates — AI Gap Analysis

## Overview

The official Portainer v3 templates (`templates.json`) contain **71 templates** across the following categories:

| Category | Count | Examples |
|---|---|---|
| Database | 10 | MySQL, PostgreSQL, Mongo, Redis, CrateDB, Elasticsearch, CockroachDB, TimescaleDB |
| Edge/IIoT | 14 | Softing EdgeConnectors, OPC Router, TOSIBOX, EMQX MQTT, Mosquitto, Node-RED, Litmus Edge |
| Web/CMS | 8 | Nginx, Caddy, WordPress, Drupal, Joomla, Ghost, Plone |
| DevOps/CI | 5 | Jenkins, GitLab CE, Dokku, Registry |
| Monitoring | 4 | Grafana, Datadog, Sematext, Swarm Monitoring |
| Messaging | 1 | RabbitMQ |
| Storage | 3 | Minio, Scality S3, File Browser |
| Serverless | 2 | OpenFaaS, IronFunctions |
| Other | 6 | Ubuntu, NodeJS, Portainer Agent, OpenAMT, FDO, LiveSwitch |

## AI Template Count in Official Repo: **0**

There are **zero purely AI/ML-focused templates** in the current v3 template list.

### Closest to AI

- **Litmus Edge** (#70, #71) — Described as "enables industrial AI at scale" but is an OT data platform, not an AI deployment.
- **Elasticsearch** (#13) — Used in vector search / RAG pipelines but is a general-purpose search engine.

---

## v2 Coverage Map

This repository now provides **26 AI templates** organized into 9 sub-categories, mapped against the 4 AI infrastructure positioning pillars:

### Pillar 1: GPU-Aware Fleet Management
| Template | What It Proves |
|---|---|
| NVIDIA Triton | Multi-framework model serving across GPU fleet with dynamic batching |
| vLLM | High-throughput LLM inference with tensor parallelism across GPUs |
| NVIDIA NIM | Enterprise-grade NVIDIA-optimized inference microservices |
| Ray Cluster | Distributed GPU scheduling across head + worker nodes |
| Ollama / LocalAI | Single-node GPU inference engines |

### Pillar 2: Model Lifecycle Governance
| Template | What It Proves |
|---|---|
| MLflow + MinIO (Prod) | Versioned model registry + S3 artifact store + PostgreSQL tracking |
| Prefect | Governed pipeline orchestration with scheduling, retries, audit logs |
| BentoML | Model packaging with versioning and metrics endpoints |
| Label Studio | Data labeling with project-level access control |
| MLflow (standalone) | Experiment tracking and model comparison |

### Pillar 3: Edge AI Deployment
| Template | What It Proves |
|---|---|
| ONNX Runtime (edge profile) | CPU-only inference with memory/CPU limits for constrained devices |
| NVIDIA Triton | Supports Jetson via multiarch images, model polling for OTA updates |
| NVIDIA DeepStream | Video analytics pipeline for factory-floor cameras |

### Pillar 4: Self-Service AI Stacks
| Template | What It Proves |
|---|---|
| Open WebUI + Ollama | One-click ChatGPT-like deployment, no CLI needed |
| Langflow / Flowise | Visual drag-and-drop agent builders |
| n8n (AI-Enabled) | Workflow automation with AI nodes, accessible to non-developers |
| Jupyter GPU | Notebook environment for data science teams |

### Architecture Diagram Workloads
| Diagram Node | Template(s) |
|---|---|
| LLM Fine-Tune | Ray Cluster |
| RAG Pipeline | Qdrant + ChromaDB + Weaviate + Langflow/Flowise |
| Vision Model | DeepStream, ComfyUI, Stable Diffusion WebUI |
| Anomaly Detection | DeepStream (video analytics), Triton (custom ONNX/TensorRT models) |

---

## Remaining Gaps (Future Work)

| Gap | Why It Matters | Potential Addition |
|---|---|---|
| ARM/Jetson-native images | True edge AI on embedded devices | Triton Jetson images, ONNX Runtime ARM builds |
| Air-gapped deployment | Industrial environments with no internet | Offline model bundling scripts |
| Model A/B testing | Production model governance | Seldon Core or custom Envoy routing |
| Federated learning | Privacy-preserving distributed training | NVIDIA FLARE or Flower |
| LLM evaluation/guardrails | Safety and quality governance | Ragas, DeepEval, NVIDIA NeMo Guardrails |

---

*Generated: March 2026 — For use with Portainer Business Edition and Community Edition*