Local RAG agent setup is more complicated than expected

I messed around with Akintu, a local RAG agent, over the weekend and hit quite a few pitfalls. They call it an agent, but at its core, it's still Retrieval-Augmented Generation (RAG)—basically having the LLM retrieve relevant snippets from a private knowledge base first, then generate answers based on those snippets. This avoids retraining the entire model, which is way cheaper. Their main selling points are local execution plus knowledge graphs. Someone on Hacker News asked: once an agent interacts with a knowledge graph, does its reasoning code capability actually improve? My conclusion after testing: it depends on how you define 'improve.'