Encapsulating Textual Contents into a MOC data Structure for Advanced Applications

TL;DR

This work investigates extending the IVOA MOC standard by encoding textual descriptions (Textual MOCs) and their semantic embeddings (Semantic MOCs) within sky-region maps. It outlines JSON-based serialization for Textual MOCs, introduces embedding-based Semantic MOCs via LangChain/OllamaEmbeddings, and demonstrates integration with MocServer and visualization tools like Aladin Lite and HiPS. The approach enables advanced retrieval via RAG, multimodal AI, and context-rich queries, linking spatial footprints with text, metadata, and image data. The results highlight enhanced data exploration, automated cross-matching, and potential for agent-based, multi-messenger astronomy workflows, while outlining future directions toward VOResource alignment and scalable, AI-assisted analysis.

Abstract

Context. The Multi-Order Coverage map (MOC) is a widely adopted standard promoted by the International Virtual Observatory Alliance (IVOA) to support data sharing and interoperability within the Virtual Observatory (VO) ecosystem. This hierarchical data structure efficiently encodes and visualizes irregularly shaped regions of the sky, enabling applications such as cross-matching large astronomical catalogs. Aims. This study aims to explore potential enhancements to the MOC data structure by encapsulating textual descriptions and semantic embeddings into sky regions. Specifically, we introduce "Textual MOCs", in which textual content is encapsulated, and "Semantic MOCs" that transform textual content into semantic embeddings. These enhancements are designed to enable advanced operations such as similarity searches and complex queries and to integrate with generative artificial intelligence (GenAI) tools. Method. We experimented with Textual MOCs by annotating detailed descriptions directly into the MOC sky regions, enriching the maps with contextual information suitable for interactive learning tools. For Semantic MOCs, we converted the textual content into semantic embeddings, numerical representations capturing textual meanings in multidimensional spaces, and stored them in high-dimensional vector databases optimized for efficient retrieval. Results. The implementation of Textual MOCs enhances user engagement by providing meaningful descriptions within sky regions. Semantic MOCs enable sophisticated query capabilities, such as similarity-based searches and context-aware data retrieval. Integration with multimodal generative AI systems allows for more accurate and contextually relevant interactions supporting both spatial, semantic and visual operations for advancing astronomical data analysis capabilities.

Figures (4)

• Figure 1: An illustration showing how a Multi-Order Coverage (MOC) map can be enriched with textual content retrieved from MOCServer, allowing the simultaneous visualization of both spatial and textual components. The MOC for the dataset CDS/P/Euclid/Q1/color (Euclid Quick Data Release 1) from the Euclid space survey mission is displayed, with metadata (specifically the obs_description field) used to generate a textual MOC. The interactive visualization is provided by ipyaladin, a Jupyter/IPython widget, showing both MOCs and their corresponding image datasets in HiPS format.
• Figure 2: An illustrative example demonstrating the integration of the Textual MOC data structure into a web page using Aladin Lite. The figure illustrates the Aladin Lite viewer with the MOC map, an informative popup activated by the cursor, and a menu for managing sound effects. Interactive demo: https://ggreco77.github.io/TextualMOC/AladinGame/aladin_game_d1/
• Figure 3: An interactive game using Textual MOCs within Aladin Lite helps users locate the Large Magellanic Cloud (LMC). The magenta-colored MOC marks the position of the LMC, and informational popups provide feedback as the user interacts with the map. Auditory cues enhance the user experience by signaling correct or incorrect selections. Interactive demo: https://ggreco77.github.io/TextualMOC/AladinGame/aladin_game_d2/
• Figure 4: RAG Processing Framework with spatial MOC integration and vision analysis. From a user query, (1) we embed textual MOCs into vector representations and (2) index them in a vector database. In (3) a retrieval step returns a set of similarity scores ($\text{sim}_k$) and their corresponding record indices ($\text{idx}_k$), where each $\text{sim}_k$ quantifies the relevance of the $k$-th MOC to the query and $\text{idx}_k$ points to the original MOC entry. These indices allow us to (4) fetch the matching spatial MOC objects, (5) generate context-grounded answers with a language model, (6) overlay the selected MOCs on all-sky imagery via Aladin Lite, and (7) stream the FITS image URLs (image) to a vision LLM for automated visual feature analysis.