Developer Tests Markdown and Knowledge Graphs for Claude CLI ERP Queries
A Reddit user described an experimental Claude CLI architecture for a complex ERP application that combines structured Markdown documentation, generated knowledge graphs, and read-only database access through a custom MCP server. The developer reported better handling of cross-module questions than with an embeddings-and-vector-search RAG setup, but estimated a 95% reduction in hallucinations without conducting a proper benchmark. The system has not been deployed in production, and multi-user data isolation remains under development.
The report offers a concrete documentation-first alternative to vector-search RAG for enterprise applications, but its performance and hallucination-reduction claims remain unverified.
I want to share something I've been experimenting with recently. Not claiming this is some revolutionary architecture or that I've solved RAG. I'm just curious if anyone else has tried something similar.
I was building an AI system for a fairly complex ERP application. Initially, I used the usual RAG approach: embeddings, vector search, retrieving relevant chunks, and passing them to the LLM.
But honestly, I wasn't happy with the results.
The problem wasn't always the LLM. Sometimes the retrieved information itself was incomplete. The model could understand individual database tables, but understanding how those tables, modules, and business rules connected was a different challenge.
So I thought, what if I stopped relying on vector search for understanding the application and gave Claude a properly structured knowledge base instead?
Here's what I did.