The pattern
A knowledge graph stores knowledge as nodes and edges: things, and named relationships between them. This site is built from one. Twenty-three small files describe this project, the concepts it rests on, the steps that build it, and where the pattern breaks.
The agent below retrieves from that graph without a vector database and without embeddings. Its entire retrieval apparatus is a flat text index: every node with its type, one line of notes, and its edges. The index sits in the system prompt. The model reads it, decides which node matters for a question, and fetches that node's file from this repository, one hop at a time, following edges by meaning rather than by similarity score.
Classic RAG asks what text is similar. This pattern asks what is connected.
The graph
Ask the agent
The agent runs in n8n. The graph index above sits in its system prompt; to answer, it may fetch node files from this repository, one hop at a time. The workflow behind it ships with the repository in two variants, alongside an MCP server workflow that serves the same graph, reads and writes, to any MCP client.