AI Docker Compose Kit

This repository provides a modular, multi-container AI environment for local development, agent orchestration, and high-performance Retrieval-Augmented Generation (RAG).

It is designed as a local AI infrastructure platform—a “local brain”—where all services operate within a unified Docker network, enabling low-latency communication, shared context, and controlled execution.

---

🧠 Architecture Overview

The system is structured into four layers:

1. Core Infrastructure

Persistent data stores and inference runtime:

• PostgreSQL, Redis, MongoDB, MySQL
• Qdrant (vector DB), SurrealDB, Infinity
• MinIO (object storage), CouchDB
• Ollama (LLM runtime)

2. Application Layer

User-facing tools and orchestration systems:

• Open WebUI, Dify, Flowise
• AnythingLLM, RAGFlow, Open-NotebookLM
• n8n (automation), Gitea (code), SearXNG (search)
• Obsidian Remote (knowledge interface)

3. Agent Layer (Optional)

• OpenClaw: system-level AI agent with access to internal services

4. Execution Layer

• Docker Sandbox (recommended): isolated execution environment for agent tools
• Enforced egress control via proxy allowlists

---

⚙️ Installation

1. Requirements

   • Docker
   • NVIDIA Container Toolkit (optional, for GPU acceleration)

2. Clone & Prepare

   git clone https://github.com/xormania/ai-docker-stack
   cd ai-docker-stack
   mkdir obsidian_vault

3. Configure Environment

   cp .env.example .env

   Edit .env and change all credentials before exposing ports.

4. Start Core + Apps

   docker compose -f compose.yaml -f compose.apps.yaml up -d

5. Optional: Add OpenClaw

   docker compose -f compose.yaml -f compose.apps.yaml -f compose.openclaw.yaml up -d

6. Optional: Dev Tools

   docker compose -f compose.yaml -f compose.apps.yaml -f compose.dev.yaml up -d

---

🚀 Startup Behavior

• Core services must pass health checks before dependents start

• ollama-init:

  • waits for Ollama
  • pulls ${OLLAMA_BOOTSTRAP_MODEL}
  • exits once complete

• All LLM-dependent services wait for model readiness

This prevents race conditions across the stack.

---

🔗 Internal Integration Matrix

All services communicate via internal Docker DNS:

• open-webui → ollama
• dify → postgres, redis, qdrant, mongodb, couchdb
• flowise → postgres, qdrant, ollama
• n8n → postgres, redis, ollama, gitea, searxng, mercure
• anythingllm → ollama, qdrant
• ragflow → mysql, redis, minio, infinity, ollama
• open-notebooklm → surrealdb, qdrant, ollama
• obsidian-remote → couchdb

---

🔐 Security Model

Internal Access

• Services communicate freely within internal Docker networks

External Egress

• Deny-by-default
• HTTPS allowed only via allowlist (recommended)

Agent Execution

• OpenClaw executes tools through Docker Sandbox
• No direct host access
• No Docker socket exposure

---

🌐 Egress Policy (Recommended)

Example allowlist:

github.com
api.github.com
raw.githubusercontent.com
pypi.org
files.pythonhosted.org
registry.npmjs.org
deb.debian.org
security.debian.org

All outbound traffic from agent execution should pass through a proxy enforcing this policy.

---

🧪 Verification

docker compose config >/tmp/stack.rendered.yml
docker compose ps
docker compose exec ollama ollama list

Quick health check:

docker compose exec n8n sh -lc \
'wget -qO- http://qdrant:6333/healthz && echo && wget -qO- http://ollama:11434/api/tags'

---

🧰 Component Breakdown

Core Engine

• Ollama — local LLM runtime
• Open WebUI — primary UI

Agentic Systems

• Dify — full-stack LLM app platform
• Flowise — visual LLM pipelines
• OpenClaw — system-level agent (optional)

Dev & Analysis

• Jupyter — experimentation
• Open-NotebookLM — document reasoning
• Gitea — local Git hosting

RAG & Search

• AnythingLLM, RAGFlow
• Unstructured API — document parsing
• SearXNG — privacy-first search

Automation & Knowledge

• n8n — workflow orchestration
• Obsidian Remote — knowledge interface
• Mercure — real-time updates

Storage

• PostgreSQL, MongoDB, Redis
• Qdrant, SurrealDB, Infinity
• MySQL, MinIO, CouchDB

---

🧠 Value Proposition: Local AI Infrastructure

This stack provides:

Zero-Latency RAG

Internal networking eliminates external API overhead.

Privacy-First Operation

All data, inference, and orchestration remain local.

Unified Context

Every service shares access to:

• obsidian_vault
• local repositories (Gitea)
• internal vector memory (Qdrant)

Agent-Orchestrated Workflows

With OpenClaw + n8n:

• trigger workflows from notes
• perform autonomous reasoning
• update codebases
• query local + external data (controlled)

---

⚠️ Design Principles

• No Docker socket exposure
• No privileged containers
• Internal-first networking
• Explicit egress control
• Composable architecture via multiple compose files

---

🧭 Future Direction

• Agent API layer (OpenClaw integration)
• Dynamic egress policies
• Model routing and optimization
• Dataset versioning via MinIO

---

🧩 Summary

This repository is not just a collection of tools.

It is a local AI operating environment:

• inference
• memory
• automation
• orchestration
• development

All running within a controlled, extensible Docker system.

---