Bidirectional Semantic Complementary Tool Retrieval for Remote Sensing Agents

Researchers propose a bidirectional semantic complementary tool retrieval (BSCTR) method to improve how LLM-based agents select appropriate tools for remote sensing tasks. The approach addresses a fundamental mismatch between high-level user queries and detailed tool documentation by enhancing queries with decomposed subtasks and enriching tool descriptions with contextual dependencies, demonstrating improved performance on specialized remote sensing benchmarks.

This research tackles a critical bottleneck in agentic AI systems: the ability to retrieve the right specialized tools from large libraries when context windows are limited. As LLM agents become more prevalent in domain-specific applications like remote sensing, the challenge of connecting abstract user intentions to granular technical tools has become increasingly important. The paper identifies and addresses 'semantic asymmetry'—the fundamental mismatch where users express general goals while tools provide technical specifications that lack workflow context.

The proposed solution operates bidirectionally. On the query side, agents decompose user intentions into logical subtasks, providing explicit functional semantics that better match tool capabilities. On the tool side, the researchers construct a dynamic dependency graph that captures how remote sensing tools interact in sequences, ensuring that tool representations include contextual information from prerequisite operations. This approach recognizes that specialized domains like remote sensing involve tightly coupled tool chains where understanding dependencies is essential.

The methodology demonstrates broad applicability beyond remote sensing. Testing on both GeoPlan-bench (specialized remote sensing dataset) and API-Bank (general-purpose API dataset) shows the approach generalizes effectively. For developers building agentic systems, improved tool retrieval directly enhances reliability and reduces unnecessary API calls. For AI researchers, this work provides practical techniques for bridging semantic gaps in specialized domains where tool ecosystems are complex and documentation is extensive. The open-sourcing of code and datasets enables rapid adoption and further research in agentic workflow optimization.

• →Bidirectional semantic enhancement significantly improves tool retrieval accuracy in LLM-based agents handling remote sensing tasks.
• →Planning-based query decomposition and dynamic tool dependency graphs address the core semantic asymmetry problem in agent tool selection.
• →The method demonstrates robust transfer capabilities from specialized domains to general-purpose API retrieval tasks.
• →Tightly coupled tool chains in remote sensing require contextual embeddings that incorporate predecessor tool information.
• →Open-source release of GeoPlan-bench dataset enables future research on domain-specific agentic tool retrieval.