NASA's agentic search system 5x better at finding Earth data

NASA's vast archive of geoscience datasets and tools—including Worldview, Giovanni, and the Science Discovery Engine—has long been a challenge even for domain experts. A new paper from university and NASA-affiliated researchers introduces an agentic search system that lets users pose natural-language research queries and instantly retrieve the matching datasets and workflows. The system leverages the NASA Earth Observation Knowledge Graph (NASA EO-KG) to structure metadata, then adds a neural scorer trained on NASA-EO-Bench, a new benchmark of 47,000 query-dataset pairs (21,000 task-based queries). The neural scorer alone beats cosine similarity and BM25 baselines; when fused with BM25 via score fusion, Recall@10 and Mean Reciprocal Rank (MRR) jump by more than 5x.

On top of this supervised pipeline, the team adds a zero-shot agentic reranking stage that uses LLM reasoning without any additional training. On a stratified subset of 200 queries, this step lifts MRR by 28%, demonstrating that large language models can complement traditional retrieval in a complementary way. The system is already deployed as a public-facing service for the geoscience community. The paper, submitted to arXiv, argues that knowledge graphs regain latent value when paired with agentic search—a pattern likely to extend beyond Earth observation to other scientific domains.