WikiChat

Why we included this project

WikiChat tackles one of the most practical pain points when teams wire LLMs to real data: confident but wrong answers. It wraps a retrieval step around a conversational model so each reply is grounded in passages pulled from a corpus, with citations attached. For developers building fact-heavy assistants, say an internal helpdesk that answers from company documentation or a research copilot that must stay close to its sources, this offers a ready-made pipeline you can run from the terminal and adapt to your own index instead of only Wikipedia. The project also ships a distillation path to run a smaller, cheaper model for lower-latency deployment. If hallucination is what keeps your RAG prototype from going to production, this is a well-documented starting point worth reading before you design your own retrieval loop.