WikiMind: how we built a RAG for an AI chat over corporate documents

While everyone is debating AI agents and whether an LLM will replace knowledge base search, we at Gainad took the practical route: we built WikiMind — a corporate AI assistant on a RAG architecture.

Here’s how it works: we split documents from the Wiki into chunks, turn them into embeddings, and store them in a vector database. When a user asks a question, the system looks for the most relevant passages and adds them to the LLM context. As a result, the model doesn’t make things up “as if it remembers” — it answers based on your data, which is a real must-have for an enterprise.

We built 4 chat modes: corporate AI assistant, ChatGPT, HR, and FD. Each one has its own knowledge base, and everything is tied to a role-based access model. In other words, HR doesn’t get into finance, and finance doesn’t get where it shouldn’t. Just like in the office: coffee is shared, access rights are not 🙂

At its core, this is about how to connect an LLM with internal documents without chaos, leaks, and the classic “oops, it answered the wrong thing.”