Telepix Open-Sources 'Pixy,' a Satellite Data AI Search

Accelerating Space AI Innovation: TelePIX Unveils Specialized AI Search Model 'PIXIE' as Open Source TelePIX, a leader in space AI solutions, is driving AI technology innovation in the space industry by fully releasing its 'PIXIE' series, an AI search model specialized in satellite and space technical documents, as open source. This groundbreaking model plays a critical role in finding the most accurate and reliable information from vast technical documents before a large language model (LLM) generates an answer. While existing general-purpose search models have shown limitations in understanding complex terminology and contexts within highly specialized technical domains such as aerospace, satellites, and defense, PIXIE was custom-designed to clearly address these issues. Developed as an embedding model optimized for specialized technical documents, it maximizes information accuracy and interpretability. The PIXIE series consists of four powerful models: * **PIXIE-Splade:** A Korean-exclusive sparse vector model, excellent for quickly finding documents containing specific keywords within large document sets. * **PIXIE-Rune:** An encoder-based embedding model proficient in Korean and English, optimized for semantic search that accurately grasps context and meaning. * **PIXIE-Spell:** A decoder-based embedding model supporting multilingual search, offered in two versions with 0.6 billion (0.6B) and 1.7 billion (1.7B) parameters to meet diverse requirements. All these PIXIE models are already applied in TelePIX's satellite image analysis AI agent solution, 'SatChat,' playing a critical role in accurately understanding user query intent and searching for relevant information. TelePIX stated that PIXIE has demonstrated excellent performance by recording high NDCG (Normalized Discounted Cumulative Gain) scores in Korean and English search benchmarks, placing search results that align with user intent at the top. The open-source release of the PIXIE series was carried out via the AI model sharing platform Hugging Face. This marks the first instance among domestic space companies to release a self-developed AI embedding model as open source, making its significance considerable. TelePIX expects this to accelerate the overall advancement of AI technology in the space industry and significantly contribute to the expansion of the space AI ecosystem through open collaboration. Park Jae-wan, Head of TelePIX's AI Research Team, emphasized, "The search performance of the PIXIE model is a critical factor determining the quality of responses generated by SatChat," adding that the open-source release of the PIXIE series will foster free model utilization by researchers and companies and accelerate the pace of technological development across the industry. Meanwhile, TelePIX's SatChat is an AI agent trained on satellite-based space big data, offering various functionalities such as generating answers to natural language queries and displaying maps. This innovative solution recently garnered global recognition for its technological prowess, successfully being exported to a Polish satellite company along with satellite imagery from its CubeSat 'BlueBon'.