RAG That Actually Helps, Not One That “Hallucinates in Prod”
At Gainad, we built WikiMind — an AI chat over internal corporate documents. Documents are split into chunks, turned into embeddings, and the most relevant fragments are retrieved for the user’s question and added to the LLM context.
At Gainad, we built WikiMind — an AI chat over corporate documents. The setup is classic, but the devil is in the details: we take documents from the Wiki, split them into chunks, turn them into embeddings, and store them in a vector database. Then, for each user question, we look for the most relevant fragments and add them to the LLM context.
The key point is that this is not one “smart bot for everything.” We have 4 chat modes: corporate AI assistant, ChatGPT, HR, and FD. Each one has its own knowledge base and access controlled by a role model. In other words, HR does not get into finance, and FD does not answer Wiki questions “from memory.”
Against the backdrop of the hype around AI agents and enterprise search, this is basically a must-have: less manual searching, fewer “where is this policy stored?” questions, more answers that actually matter. In short, we built it so AI inside the company is not a meme, but a working tool.