Space artificial intelligence (AI) solutions corporations TelePIX said on the 3rd that its aerospace technical document search AI model "PIXIE-v1.0" ranked second worldwide in the Retrieval Embedding Benchmark (RTEB) category for models with 1 billion (1B) parameters or fewer.
PIXIE-v1.0 is an aerospace domain–specialized information retrieval model designed to enable semantic search for highly specialized technical documents in aerospace, satellites, and defense. It was designed to make it easier to search massive aerospace technical documents—such as satellite design documents, technical specifications, and operation manuals—using natural-language queries. Developed to enhance the performance of the satellite agent AI solution SatCHAT, it follows the earlier release of PIXIE-Preview and in version 1.0 focuses on quantitatively verifying domain-specialized search performance.
Models competing with PIXIE-v1.0 included numerous embedding models released by global big tech and research institutions. RTEB is a next-generation search benchmark that extends the Massive Text Embedding Benchmark (MTEB) leaderboard, which has been used as the standard for evaluating embedding models. It focuses on assessing AI model information retrieval performance in real industrial environments rather than competition centered on test data scores. It can verify the model's practical industrial applicability based on high-difficulty domains such as law, finance, medicine, and code.
While most, including the top model in the category, are general-purpose models spanning multiple domains such as law, finance, medicine, and code, TelePIX's PIXIE-v1.0 achieved a top global result even though it focused on the aerospace domain and Korean-English technical documents, the company said.
The company said the result shows that high search performance can be achieved through domain-specialized data curation and training quality improvements rather than scaling up model size. It noted that semantic search operated stably even in aerospace document environments where specialized terms and abbreviations are intricately intertwined, confirming applicability on industrial sites.
TelePIX also conducted an additional evaluation using its in-house search benchmark "STELLA" to verify multilingual aerospace domain search performance, including Korean, which RTEB does not directly address. As a result, PIXIE-v1.0 showed high search accuracy relative to its parameter size and secured stable language- and domain-specialized search capabilities, TelePIX said. STELLA is a multilingual information retrieval benchmark based on specialized aerospace documents, designed to compensate for the practical limitation that there are almost no public search evaluation criteria specialized for the aerospace domain.
TelePIX has open-sourced PIXIE-v1.0 and expects it can be used as a core model for specialized technical document search in Retrieval-Augmented Generation (RAG)–based AI systems.
Kwon Darongsae, head of the TelePIX data science institutional sector, said, "The PIXIE-v1.0 we released this time maintained the direction presented in the preview stage, while focusing on more stably advancing aerospace domain search performance to achieve excellent results in performance evaluations," and added, "We expect PIXIE and STELLA to serve as foundational resources for future research on domain-specialized information retrieval and real-world applications."