p4n4-ai

Dockerized GenAI stack — local LLM inference, stateful AI agents, and workflow automation.

The GenAI stack (Ollama · Letta · n8n) brings local AI capabilities to your IoT deployment. Ollama runs open-weight LLMs entirely on-device, Letta provides persistent AI agents with long-term memory, and n8n wires everything together with event-driven workflows.

Attaches to the shared p4n4-net Docker bridge network created by p4n4-iot, enabling seamless integration with MQTT, InfluxDB, Node-RED, and Grafana.

Part of the p4n4 platform — an EdgeAI + GenAI integration platform for IoT deployments.

---

Table of Contents

• Architecture
• Stack Components
• Prerequisites
• Getting Started
• Project Structure
• Ollama Models
• n8n Workflows
• GPU Support
• Usage
• Default Ports
• Default Credentials
• Network Requirements
• Security Hardening
• Local Overrides
• Integration with p4n4-iot
• Resources
• License

---

Architecture

  [p4n4-iot / MQTT / InfluxDB]
           │
           │  (shared p4n4-net bridge)
           ▼
        [n8n]           ← event-driven workflow automation
       /     \
      ▼       ▼
  [Ollama]  [Letta]     ← local LLM runtime + stateful AI agents

Data flow: n8n subscribes to MQTT topics (via p4n4-net) and triggers AI workflows. Ollama serves local LLM inference, Letta manages persistent AI agents with memory, and n8n routes results back to MQTT, InfluxDB, or external webhooks.

---

Stack Components

Service	Role	Description
Ollama	Local LLM Runtime	Runs open-weight models (Llama, Mistral, Phi, etc.) entirely on-device with zero data egress. Exposes an OpenAI-compatible REST API on port 11434.
Letta	AI Agent Framework	Stateful AI agent framework with persistent memory (formerly MemGPT). Build agents that remember context across sessions and reason over long-term IoT event histories.
n8n	Workflow Automation	Low-code, node-based workflow engine. Connects MQTT, InfluxDB, Ollama, Letta, and external APIs without custom glue code. Includes four starter IoT + AI workflows.

---

Prerequisites

• Docker (v20.10+)
• Docker Compose (v2.0+)
• At least 8 GB RAM available to Docker (16 GB recommended for larger models)
• p4n4-iot running (or p4n4-net network created manually — see Network Requirements)
• (Optional) NVIDIA GPU with drivers + NVIDIA Container Toolkit for GPU acceleration

---

Getting Started

1. Clone the repository

   git clone https://github.com/raisga/p4n4-ai.git
   cd p4n4-ai

2. Configure environment variables

   cp .env.example .env
   # Edit .env — at minimum change N8N_ENCRYPTION_KEY and passwords

3. Ensure p4n4-net exists (skip if p4n4-iot is already running)

   docker network create p4n4-net

4. Start the stack

   docker compose up -d
   # or
   make up

5. Pull a language model

   make pull-models
   # or pull specific models:
   ./scripts/pull-models.sh llama3.2 nomic-embed-text

6. Open the interfaces

   • n8n: http://localhost:5678
   • Letta: http://localhost:8283
   • Ollama API: http://localhost:11434

---

Project Structure

p4n4-ai/
├── docker-compose.yml                  # GenAI stack service definitions
├── docker-compose.override.yml.example # Local override template (GPU, dev)
├── Makefile                            # Convenience commands
├── .env.example                        # Environment template (copy to .env)
├── .gitignore
├── config/
│   ├── ollama/                         # Ollama config (models pulled at runtime)
│   ├── letta/
│   │   └── letta.conf                  # Letta server configuration reference
│   └── n8n/
│       └── workflows/
│           ├── alert-enrichment.json   # Enrich MQTT alerts with LLM analysis
│           ├── scheduled-digest.json   # Hourly telemetry summary via Ollama
│           ├── device-onboarding.json  # Auto-register new MQTT devices
│           └── incident-escalation.json # Classify and escalate critical alerts
└── scripts/
    ├── pull-models.sh                  # Helper to pull models into Ollama
    ├── selector.sh                     # Interactive service selector
    └── check_env_example.py            # CI: .env.example completeness check

---

Ollama Models

Models are not bundled in the image — pull them after starting the stack.

Pulling Models

# Pull the default model (llama3.2)
make pull-models

# Pull specific models
./scripts/pull-models.sh llama3.2
./scripts/pull-models.sh llama3.2 nomic-embed-text phi3.5

# Pull directly via Docker
docker exec p4n4-ollama ollama pull llama3.2

Recommended Models

Model	Size	Use Case
llama3.2	2 GB	General inference, alert analysis, summaries
phi3.5	2.2 GB	Lightweight reasoning, classification
nomic-embed-text	274 MB	Embeddings for Letta agent memory
llama3.2:70b	40 GB	High-quality reasoning (requires GPU)

Listing Installed Models

make models
# or
docker exec p4n4-ollama ollama list

---

n8n Workflows

Four starter workflows are included in n8n/workflows/. Import them via the n8n UI:

1. Open n8n at http://localhost:5678
2. Go to Workflows → Import from File
3. Select the JSON file from config/n8n/workflows/

Workflow	Description
alert-enrichment.json	Subscribes to inference/results MQTT topic; sends low-confidence results to Ollama for analysis
scheduled-digest.json	Runs hourly; queries InfluxDB for recent telemetry and generates a natural-language summary via Ollama
device-onboarding.json	Listens on devices/+/register; auto-registers new devices and publishes a confirmation to MQTT
incident-escalation.json	Listens on alerts/+/critical; classifies severity via Ollama and publishes enriched alert to alerts/escalated

MQTT Credential Setup in n8n

After importing workflows, configure an MQTT credential named p4n4 MQTT:

• Host: p4n4-mqtt (service name on p4n4-net)
• Port: 1883
• Username/Password: match values in your p4n4-iot .env

---

GPU Support

To enable NVIDIA GPU acceleration for Ollama, use the override file:

cp docker-compose.override.yml.example docker-compose.override.yml
# Uncomment the 'ollama' GPU section
docker compose up -d

Verify GPU detection:

docker exec p4n4-ollama nvidia-smi

For AMD (ROCm), uncomment the ollama:rocm override section instead.

---

Usage

Make Commands

make help             # Show all available commands

make up               # Start the full stack
make down             # Stop all services
make restart          # Restart all services
make logs             # Follow logs from all services
make ps               # Show service status
make status           # Colorized status table

make start SERVICE=n8n    # Start a single service
make stop SERVICE=letta   # Stop a single service

make models           # List Ollama models
make pull-models      # Pull default models
make test-ollama      # Send a test prompt to Ollama

make clean            # Stop services and remove all data volumes

Testing Ollama

# Via make
make test-ollama

# Via curl (from host)
curl http://localhost:11434/api/generate \
  -H "Content-Type: application/json" \
  -d '{"model":"llama3.2","prompt":"Hello!","stream":false}'

# From another container on p4n4-net
docker run --rm --network p4n4-net curlimages/curl \
  curl -s http://p4n4-ollama:11434/api/generate \
  -H "Content-Type: application/json" \
  -d '{"model":"llama3.2","prompt":"Hello!","stream":false}'

---

Default Ports

Service	Port	URL
Ollama API	11434	http://localhost:11434
Letta Server	8283	http://localhost:8283
n8n UI	5678	http://localhost:5678

---

Default Credentials

All credentials are set in .env. Defaults from .env.example:

Service	Username	Password
n8n	admin	adminpassword
Letta	(no username)	lettapassword

Note: Change all passwords and the N8N_ENCRYPTION_KEY before deploying to production.

---

Network Requirements

This stack attaches to p4n4-net as an external network. The network must exist before running docker compose up.

Option 1 — Use p4n4-iot (recommended):

# In p4n4-iot directory
docker compose up -d
# Then start p4n4-ai

Option 2 — Create network manually:

docker network create p4n4-net
docker compose up -d

Option 3 — Use the CLI:

p4n4 up --all   # starts all stacks in the correct order

---

Security Hardening

1. Change all default credentials in .env before exposing services externally.

2. Set a strong N8N_ENCRYPTION_KEY — this encrypts stored credentials in n8n. Minimum 32 characters.

3. Letta API password — set LETTA_SERVER_PASSWORD to a strong value. All API calls require this password as a bearer token.

4. Restrict port exposure — for production, remove host-port bindings from docker-compose.yml and access services only via p4n4-net or a reverse proxy.

5. Ollama access — Ollama has no built-in authentication. Restrict access at the network or reverse-proxy level for production deployments.

---

Local Overrides

Use docker-compose.override.yml for machine-specific settings (GPU, external hostnames, custom volumes):

cp docker-compose.override.yml.example docker-compose.override.yml
# Edit docker-compose.override.yml as needed
docker compose up -d

The override file is listed in .gitignore and will never be committed.

---

Integration with p4n4-iot

When running alongside p4n4-iot on the same p4n4-net network, services can be referenced by their container names:

p4n4-iot Service	Address from p4n4-ai
MQTT Broker	p4n4-mqtt:1883
InfluxDB	p4n4-influxdb:8086
Node-RED	p4n4-node-red:1880

Use these addresses in n8n workflow nodes, Letta agent configurations, and Ollama-powered scripts.

Shared secrets (must match between stacks — set identical values in both .env files):

Variable	Purpose
INFLUXDB_TOKEN	InfluxDB API token
INFLUXDB_ORG	InfluxDB organization
INFLUXDB_BUCKET	Primary InfluxDB bucket

---

Resources

• p4n4 Platform — umbrella repo and architecture docs
• p4n4-iot — IoT stack (MING)
• Ollama Documentation — available models and API reference
• Letta Documentation — agent framework docs
• n8n Documentation — workflow automation docs

---

License

This project is licensed under the MIT License.