NestBot is a natural language interface for NEST, Empa's modular research and innovation building in Dübendorf. NEST hosts over 10'000 datapoints alongside student projects, publications, building documentation in the form of architectural drawings and BIM. This data that is not easy to find and currently accessible only to researchers with technical expertise in SQL, REST APIs, or specialised tools. This project developed NestBot, a Retrieval-Augmented Generation (RAG) chatbot that allows any researcher or building manager to query NEST data in plain language. A user can ask "Which sensors measure CO₂ on the second floor of UMAR?" and receive a direct, grounded answer. Alongside NestBot, we developed AutoRAG: a semi-automatic framework for evaluating and continuously improving RAG systems. AutoRAG generates synthetic test questions from the knowledge base, evaluates system performance, diagnoses failure patterns through LLM-based root cause analysis, and guides targeted improvements. Over three improvement iterations, NestBot's retrieval success rate improved from 30% to 66%. A key finding from the project — emerging directly from the evaluation work and a practical experiment connecting Claude as an agent to the NEST database — has led to a strategic reorientation: the next phase will move from a rigid RAG pipeline to an agent-based architecture, enabling more flexible, self-correcting data access. Both NestBot and AutoRAG are designed as open, reusable tools applicable beyond this project.