Air-Gapped Local n8n RAG Suite: Dual-Tool AI Copilot & Vector Sync Engines
Standard cloud LLMs frequently hallucinate n8n node schemas, parameters, and expression syntax ($json vs $input). Plus, sending internal workflow logic or credentials to SaaS models creates compliance risks.
I built the n8n Local RAG Suite: an end-to-end, fully air-gapped retrieval-augmented generation architecture running 100% locally on host hardware (Ollama + Qdrant + n8n).
Architecture Overview
[ Doc Ingestion ] [ Template Sync ]
(Daily 3 AM Sync) (749k Catalog)
β β
βΌ βΌ
[ Qdrant: n8n_docs ] [ Qdrant: n8n_templates ]
(14.6k Vectors) (749k Vectors)
β β
ββββββββββββ¬ββββββββββββ
β
βΌ
[ AI Engineering Agent ]
(qwen2.5:7b-instruct)
Key Highlights
- ~10ms Smart GitHub SHA Guard: Checks the latest
n8n-io/n8n-docscommit SHA against stored Qdrant memory, skipping unnecessary re-indexing if documentation hasnβt changed. - Pre-Ingestion Secret Sanitizer: Runs 7 regex security filters to scrub raw credentials, API keys, and emojis before vector storage.
- Dual-Tool Local RAG Routing:
qwen2.5:7b-instructdynamically routes between official documentation (n8n_docs_retriever) and community workflow patterns (n8n_templates_retriever).
Hardware Requirements
- Local Hardware: Runs on workstation GPUs (tested on RTX 3000 Ada / 8GB VRAM).
- Ollama Models:
qwen2.5:7b-instruct(num_ctx: 16384) +nomic-embed-text. - Vector DB: Qdrant (
qdrant/qdrant:latest).
Get Started & Full Documentation
Full installation steps, Docker Compose files, metadata schemas, and test cases are available on GitHub. You can also skip the multi-hour template embedding sweep by downloading the pre-indexed 749k vector snapshot directly from Hugging Face.
GitHub Repository: jdm6457/n8n-rag-suite
Instant Qdrant Database Seed: Hugging Face Dataset Snapshot
Would love to hear your feedback or ideas for the planned nine (9) additional phases!