Table of Contents

8 Sections

BLUEPRINT OVERVIEW

Overview

This blueprint details how to build a 100% private, on-premise Retrieval-Augmented Generation (RAG) agent. It connects a local Ollama instance running DeepSeek-R1 (or Llama 3.3) with a self-hosted Qdrant vector database and n8n orchestration. No customer data or internal company documents ever leave your private server network.

Before you begin

This blueprint is autonomously synthesized and validated against Workflow Vault's Astro Content Collections schema. Review prerequisites and test against sandbox credentials prior to production deployment.

TERMINOLOGY

Glossary

RAG (Retrieval-Augmented Generation): An architecture where an LLM is provided with relevant factual context retrieved from a private database before generating an answer, preventing hallucinations.
DeepSeek-R1: An open-weights reasoning model that uses reinforcement learning to produce verifiable, step-by-step chain-of-thought logic before answering.
Qdrant: An open-source vector similarity search engine written in Rust, optimized for fast semantic filtering and payload storage.
Ollama: A lightweight CLI and API server that downloads, optimizes, and runs open-source language models locally on your GPU or CPU.

REQUIREMENTS

Prerequisites

[You'll need this already] A Linux server (VPS or local workstation) with at least 16 GB RAM (NVIDIA GPU recommended for sub-second inference, or modern CPU for 8B quantized models).
[You'll need this already] Docker Engine 24+ and Docker Compose v2 installed.
[You'll need this already] Basic familiarity with running shell commands over SSH.
[Optional, not required] Cloudflare Tunnel or Caddy if exposing the n8n webhook endpoint to external Slack/Discord bots.

RUNTIME ENVIRONMENT

Where to Run This Automation

Where should you run this automation?

Don't have 24/7 hardware at home? Choose the setup that fits your budget and technical comfort.

Hardware & Cloud Guide

Best for Freelancers, Agencies, and 24/7 Client Pipelines

If you don't have a dedicated server running at home, a Virtual Private Server (VPS) gives you a dedicated Linux server in the cloud that stays online 24/7 even when your personal laptop is closed.

Cost: ~$4 to $6 / month (predictable flat fee, no per-execution surprises).
Providers: Hetzner Cloud, DigitalOcean, Linode / Akamai, Railway, or OVH.
Key advantage: You get a permanent public IP address and free SSL certificates, which are essential for receiving incoming Stripe, Airtable, or GitHub webhooks reliably.

Quick 1-liner to install Docker on any Ubuntu/Debian VPS: curl -fsSL https://get.docker.com | sh && sudo usermod -aG docker $USER

Best for Hobbyists, Testing & Zero-Cost Experimentation

You can run this automation completely free on your everyday computer (Mac, Windows, or Linux) or on low-cost home hardware like a Raspberry Pi 4/5 or repurposed mini PC.

Cost: $0 / month (100% free).
Software needed: Docker Desktop (free for personal use) or native Node.js / Python.
Consideration: Automations will only execute while your computer is powered on and awake. To receive incoming webhooks locally, use tools like ngrok or Cloudflare Tunnels for secure local tunneling.

Best for Zero-Maintenance 1-Click Operations

If you don't want to manage Docker containers or terminal commands, you can run automations directly on managed serverless platforms or official SaaS plans:

Cloudflare Workers: Free tier includes 100,000 requests per day with 0ms cold starts.
n8n Cloud / Make / Zapier: Fully managed hosted platforms with automated backups and visual scenario builders.
Railway / Render: 1-click Git container deploys with auto-healing and managed PostgreSQL databases.

EXECUTION PLAYBOOK

Step-by-step implementation

Click the numbered phase buttons as you complete each step to track your live progress.

Phase Progress 0/3 completed (0%)

PHASE 1

Step 1: Define the Unified Multi-Container Compose Stack

Create a dedicated directory `ai-rag-stack` and define `docker-compose.yml` declaring n8n, Ollama with GPU acceleration (or CPU fallback), and Qdrant storage.

yaml

services:
  ollama:
    image: ollama/ollama:latest
    container_name: ollama_llm
    restart: unless-stopped
    volumes:
      - ollama_data:/root/.ollama
    ports:
      - "127.0.0.1:11434:11434"

  qdrant:
    image: qdrant/qdrant:latest
    container_name: qdrant_db
    restart: unless-stopped
    ports:
      - "127.0.0.1:6333:6333"
    volumes:
      - qdrant_data:/qdrant/storage

  n8n:
    image: n8nio/n8n:latest
    container_name: n8n_rag_engine
    restart: unless-stopped
    ports:
      - "127.0.0.1:5678:5678"
    environment:
      - N8N_HOST=localhost
      - N8N_PORT=5678
      - N8N_PROTOCOL=http
      - WEBHOOK_URL=http://localhost:5678/
    volumes:
      - n8n_data:/home/node/.n8n
    depends_on:
      - ollama
      - qdrant

volumes:
  ollama_data:
  qdrant_data:
  n8n_data:

Why this matters

Running n8n, Ollama, and Qdrant in the same Docker network provides high-speed internal socket communication and eliminates public exposure of the database.

Expected Output

All 3 containers start cleanly without port conflicts.

PHASE 2

Step 2: Pull the DeepSeek-R1 Model & Embedding Model

Execute the Ollama CLI inside the running container to pull the quantized reasoning model and a high-speed embedding model.

bash

# Pull DeepSeek-R1 (8B distilled reasoning model)
docker exec -it ollama_llm ollama run deepseek-r1:8b

# Pull high-performance BGE embedding model for vector indexing
docker exec -it ollama_llm ollama pull bge-m3

Why this matters

The LLM needs weights downloaded to disk before the n8n agent node can establish inference sessions.

Expected Output

Ollama downloads model weights and responds with a success status.

PHASE 3

Step 3: Configure the n8n RAG Chain & Vector Store Connector

Open n8n at `http://localhost:5678`, add an `AI Agent` node, configure `Ollama Chat Model` (Host: `http://ollama:11434`, Model: `deepseek-r1:8b`), attach the `Qdrant Vector Store` tool, and set top-k retrieval to 4.

json

{
  "nodes": [
    {
      "name": "Ollama DeepSeek-R1",
      "type": "@n8n/n8n-nodes-langchain.lmChatOllama",
      "parameters": {
        "model": "deepseek-r1:8b",
        "baseURL": "http://ollama:11434"
      }
    },
    {
      "name": "Qdrant Vector Store",
      "type": "@n8n/n8n-nodes-langchain.vectorStoreQdrant",
      "parameters": {
        "qdrantUrl": "http://qdrant:6333",
        "collectionName": "company_docs"
      }
    }
  ]
}

Why this matters

Connecting the vector retriever node to Ollama ensures n8n queries Qdrant first, injects relevant document chunks into memory, and passes the augmented prompt to DeepSeek-R1.

Expected Output

n8n verifies the connection to both internal containers.

SYSTEM TOPOLOGY

Workflow architecture

Webhook / Slack Trigger -> n8n AI Agent Node -> Ollama (DeepSeek-R1 Reasoner) <-> Qdrant Vector DB (Dense Embeddings) -> Synthesized Response with Citations

DeepSeek-R1

n8n

Qdrant

Docker

Ollama

VERIFICATION & NEXT PHASES

Final result

A private, zero-subscription RAG pipeline processing sensitive queries locally with step-by-step reasoning and semantic document retrieval.

Recommended Extensions & Scaling

Add automated document ingestion triggers from Google Drive or local folders
Set up systemd auto-restart on host reboot

DEBUGGING & EDGE CASES

Troubleshooting

⚠ Ollama returns HTTP 500 or out-of-memory error

Root Cause

The selected model size exceeds available system RAM or VRAM.

Solution

Switch to a smaller quantized model such as deepseek-r1:1.5b or allocate swap space on Linux (`sudo fallocate -l 8G /swapfile`).

⚠ n8n cannot connect to http://ollama:11434

Root Cause

Docker containers are running on separate bridge networks or using localhost instead of container hostnames.

Solution

Use the internal Docker service name `http://ollama:11434` rather than `http://localhost:11434` inside n8n.