BROWSE
Detail
Conference Paper
출처 인지 에이전트: 어세이 컨텍스트 판단과 그래프 RAG
Cited - time in 
Share 
Share
- Authors
- 김성수
- Issue Date
- 2026-06
- Citation
- 대한전자공학회 학술 대회 (하계) 2026, pp.123-127
- Publisher
- 대한전자공학회
- Language
- Korean
- Type
- Conference Paper
- Abstract
- While large language models and retrieval-augmented gen- eration (RAG) are rapidly permeating the therapeutic AI landscape, a critical barrier persists: context fragility—the phenomenon whereby the meaning of a molecular label changes when assay semantics, endpoint definitions, deci- sion thresholds, or data provenance shift, even for the iden- tical molecule. We reframe this problem through the lens of agentic AI and propose AssayKG-RAG, a provenance- aware typed knowledge graph retrieval framework in which the agent actively plans admissible evidence, criticizes the coherence between the query context and the retrieved items, and verifies the directional validity of predictions. The system is built on three pillars: (i) canonicalization of the assay context into seven typed fields, (ii) prove- nance tags on every evidence edge, and (iii) a determinis- tic planner–critic–verifier loop that emits both predictions and inspectable audit signals. On a held-out-source synthetic benchmark and a targeted real-molecule concept-conflict stress test, AssayKG-RAG achieves a perfect Directional Correct- ness Coe!cient (DCC = 1.000), an error-detection AUROCerr = 0.777, and a Zero leakage rate (LeakageRate = 0.000). These results empirically demonstrate that the agent can surface and report context-dependent failure modes that average AUROC alone would render invisible. The core contribution is not a leaderboard refresh, but the formalization of an explicitly decoupled plan–criticize–verify loop with provenance-gated auditability into a reproducible, executable contract for therapeutic agentic AI.
- KSP Keywords
- Dependent failure, Error Detection, Failure Mode(FM), Language Models, Query Context, Stress test, Synthetic benchmark, context-dependent, data provenance, knowledge graph, leakage rate