What problem does it solve? Bag-of-words graph search fails on multi-word and mixed Chinese-English queries, and naive graph traversal returns hundreds of noisy results. This Skill bridges a large knowledge graph into an agent's memory layer with cost-aware query routing, five-level entity resolution, and graph-text fusion ranking. ## Core Features & Use Cases - Tiered Query Modes: Route requests across quick, graph (BFS), deep (full-text), and bridge (health audit) modes based on latency budget and depth needs. - Semantic Search Enhancement: Combine jieba tokenization, TF-IDF full-text indexing, and a five-level entity resolution fallback (exact → substring → tokenized → reverse → fuzzy), fusing graph and text results with graph-weighted scoring. - Graph Health Auditing: Report node/edge/type counts, connected papers and skills, and orphaned nodes via the bridge mode. - Use Case: An agent receives the query "眼动追踪" (eye tracking). The Skill tokenizes it with jieba, runs TF-IDF retrieval over source files, fuses results with graph neighbors, and returns at most 20 ranked, structured entries. ## Quick Start Ask the agent to query the knowledge graph for a known entity like "BPPV" in quick mode and return its depth-1 neighbors plus a connection health report.
