The preservation of archaeological sites is an important task in cultural heritage. Classical methods conserve archaeological objects in museums and provide restoration of archaeological sites threatened by decay. The improved digitalization provides the possibility to generate an accurate representation of archaeological sites by using laser scanners. The resulting point clouds can preserve the archaeological site and provide the possibility to view it in its digital form even if it no longer exists. Usually, the archaeological site comes with a lot of different material, which has been created over the years. This material provides information about the digitalized object, which helps to gain a deeper understanding about the presented archaeological site. This thesis presents an annotation system for a point-cloud renderer. The system allows adding annotations in the 3D space next to the part of the point cloud it belongs to. This helps to provide the additional information of the point cloud in the context it belongs to. Moreover, each annotation should present interesting information about specific annotated parts of the archaeological site to the viewer. Besides simple textual annotations, a variable amount of documents, such as images and PDFs, can be attached to each annotation to provide all kind of information. Several filtering techniques, including viewpoint-dependent priority filtering, are presented to control the visibility of the annotations. Moreover, a guidance system based on graphs is introduced to lead viewers to different points of interest, which are represented as annotations. To provide a clear connection between annotations and the annotated part of the point cloud, a point-selection method and a point-marking method are presented. To allow the connection of a large set of annotations to a single point cloud, these methods are developed in CUDA. This is done by extending existing methods, which create octrees in CUDA. The developed methods allow fast execution on the GPU while a CPU-based method is not able to handle such a large amount of point selections in real-time.