Knowledge Graphs in n8n are FINALLY Here!

Intro (who/what)

• Intro: This video shows how to add knowledge graphs to the Rag (RAG) template in N8N using Graffiti MCP, Neo4j, and OpenAI for relational querying and memory.

Key idea and value proposition

• Core idea: Combine a vector DB with a knowledge graph so the agent can traverse relationships between entities, not just retrieve individual chunks.
• Value: Improves relational reasoning and navigation (e.g., company → executives) while keeping fast, chunk-based retrieval for simple queries.

Core components and terms

• RAG pipeline with chunking into bite-sized pieces stored in a vector DB (e.g., Postgres PGVector, Quadrant, Pinecone)
• Knowledge graph to store entities/relationships
• Graffiti MCP (MCP server) to manage graph operations
• Neo4j as the graph database
• N8N (self-hosted) as the workflow orchestrator
• OpenAI as the LLM provider for entity/relationship extraction
• MCP server and MCP nodes in N8N

Two MCP nodes and their roles

• Add memory: inserts a document's extracted entities/relationships into the knowledge graph
• Search memory: queries the knowledge graph (used by the agent for relational questions)

How it fits the workflow (data flow; when to query graph vs. vector DB)

• Data flow: source data (e.g., Google Drive) → chunked → stored in vector DB; simultaneously, Graffiti MCP extracts entities/relationships to populate the knowledge graph
• When to query:
  • Use vector DB for single-entity or general overview queries
  • Use knowledge graph for relational queries (e.g., how two companies work together, who is an executive, etc.)
• Agent prompts can specify when to prefer graph vs. vector searches

Quick setup overview (prereqs and high-level steps)

• Prereqs:
  • Self-hosted N8N
  • Docker setup for Graffiti MCP server + Neo4j
  • OpenAI API key (and configurable model)
  • Access to host.docker.in networking
  • Consider firewall rules (secure, not open to everyone)
• High-level steps:
  • Clone Graffiti MCP repo and set up environment variables (OpenAI API key, Neo4j password, host URLs)
  • Run docker-compose up -d for Graffiti MCP + Neo4j
  • Verify containers with docker ps and docker logs; adjust ports if needed (e.g., 8000/8030 mapping)
  • Modify N8N docker-compose to expose host.docker.in; add required firewall rules to permit N8N → Graffiti MCP
  • Install the N8N community MCP node and configure an MCP client to talk to Graffiti MCP (host.docker.in + port)
  • Add two MCP tools in N8N: add memory and search memory
  • Test by inserting a document into the knowledge graph and performing a relational query
• Note: the setup emphasizes Graffiti for NLP-driven extraction and the MCP to expose graph tools to N8N

Demo highlights or examples mentioned

• Vector DB demo: getting an overview of a mock company from the vector store
• Knowledge graph demo: querying Dr. Tanaka and Dr. Chen via Graffiti MCP (relational search)
• End-to-end test: uploading a Google Drive file, adding it to the knowledge graph, and running the LLM-based extraction to populate entities/relationships
• Mention that graph operations can be slower/lucrative than pure vector queries due to the relational processing

Pros, cons, and guidance on when to use knowledge graphs vs. a pure vector DB

• Pros:
  • Enables relational queries and navigation between entities
  • More powerful for complex, relationship-based reasoning
  • Can search paths and edges between nodes (e.g., org charts, partnerships)
• Cons:
  • Slower and more expensive due to LLM-driven extraction and graph workloads
  • More setup complexity and maintenance (MCP, Neo4j, security)
• Guidance:
  • Use knowledge graphs when data is highly relational and you need traversals or graph-based reasoning
  • Use a pure vector DB for large-scale, fast similarity search without strong relational needs
  • For mixed workloads, run a hybrid: graph for relational queries, vector DB for fast chunk retrieval

Takeaways and potential next steps

• Takeaway: Knowledge graphs add relational power to RAG in N8N with Graffiti MCP, but require careful setup and awareness of latency/cost
• Next steps you might explore:
  • Deeper comparisons between graph-based vs. vector-based retrieval for your data
  • More advanced graph tools (e.g., edge navigation, complex queries) in MCP
  • Extensions to other data sources and LLM providers
  • further content on AI agents, RAG, and knowledge graphs integration