An experimental knowledge graph (KG) driven framework (10.26434/chemrxiv-2023-sljbt) was recently introduced to facilitate the integration of heterogeneous data types, encompassing both experimental data (mass spectrometry annotation, results from biological screening and fractionation) as well as meta-data available on the Web (such as taxonomies and metabolite databases). Although this KG efficiently encapsulates the different data structures and semantic relationships, retrieving specific information through structured or visually queries or programmatically is not trivial. To unlock the full potential of KGs for scientists of all fields, we designed and implemented a KG-based Artificial Intelligence (AI) agent that can convert natural language questions into programmatic data-mining tasks and generate adapted visualization. By leveraging the potential of emerging Large Language Models (LLMs) to understand semantic relationships encapsulated in KGs and mentioned in the questions, the KGAI-agent autonomously iterates to construct a SPARQL query – the widely-used open standard query language for knowledge graphs – of any submitted natural language question. After retrieving the necessary information from the KG, the KGAI-agent provides a preliminary interpretation of the results in natural language, along with relevant visualizations and statistics. In addition, with follow-up interactions between the LLM and the user, the AI-agent can be expert-guided to refine the initial results and interpretation. Released as an open-source tool, the KGAI-agent acts as a powerful assistant capable of exploring any KG. Accessible in over 50 human languages, it serves to democratize access to semantic databases without requiring bioinformatics expertise. Here we will describe the core principles and components of the KGAI-agent and illustrate how we can query metabolomics information from a KG to discover bioactive metabolites from a collection of 16,000 plant extracts.