Cartolabe is a web-based visualization system for large textual corpora. It has been initially designed to visualize scientific publication data, in particular the contents of the French scientific repository [HAL](https://hal.archives-ouvertes.fr/) containing publications from French public researchers. It has evolved into a more general purpose framework to display large document collections. Cartolabe visualizes a large textual corpus by projecting it as a 2D map. It relies on a flexible text analysis pipeline that uses a multidimensional projection (UMAP), but also several data analysis stages to create a usable visualization. Cartolabe is targeted towards an audience of scientists of any domain and assumes no special training on visualization and data analysis. It is meant to assist scientists explore articles or authors using a zoomable semantic map. Cartolabe is web-based yet scalable to millions of entries, and has been used to explore large corpora such as arXiv (1M entries), the French National Debate (4.3M entries), and Wikipedia (4.5M entries). See the application [cartolabe live](https://cartolabe.fr).