Toward Natural Language Search in the ALMA Science Archive Using LLM Source Descriptions
The ALMA Science Archive contains over 455,000 FITS images spanning a wide range of astronomical sources but currently lacks natural language search capabilities. We present a proof-of-concept semantic search system operating at the source level: rather than assigning descriptions at the project or observing unit level, we generate one BM25-optimized text description per unique astronomical source by aggregating proposal and publication abstracts from every project that observed that source across the archive. We prompt gpt-4o-mini to extract source type, physical properties, morphological description, and key scientific findings in a form optimized for BM25 retrieval rather than human reading. Of 54,060 unique source names identified in the archive, 10,758 had at least one associated abstract; 8,609 received descriptions and results were indexed into Elasticsearch for free-text retrieval. Comparing BM25 on LLM-extracted descriptions against BM25 on raw abstracts across 20 morphological and astrophysical queries, we find an average overlap of 2.0 out of 5 results, demonstrating that the two retrieval methods are complementary rather than redundant. We further implement an image-text bridge using pre-computed visual similarity scores (Stoehr et al. 2025): for each image without a direct source description, we assess whether its 30 most visually similar images share semantically consistent descriptions and assign a description only when they do, achieving coverage of 44% to 98% of all 455,448 images depending on the consistency threshold applied. This work represents a first step toward natural language search in the ALMA archive. The pipeline is designed for compatibility with locally-hosted models, enabling scaling to the full source catalog.