Publication
Semantically Grounded Agentic Access to the Asset Administration Shell: A Hybrid MCP Endpoint and Its Model-Size Requirements
Gerhard Sonnenberg; Daniel Porta
In: 2026 IEEE 31th International Conference on Emerging Technologies and Factory Automation. IEEE International Conference on Emerging Technologies and Factory Automation (ETFA-2026), IEEE, 2026.
Abstract
The Asset Administration Shell (AAS) digitizes industrial assets, yet its content stays largely passive: structured submodels coexist with unstructured PDF manuals that AI agents can neither retrieve nor act on. We present a Model Context Protocol (MCP) endpoint that gives large language model (LLM) agents unified read and write access to the AAS. An event-driven ingestion plugin synchronizes the endpoint's graph and vector stores with the AAS repositories, validated on a real-world AAS corpus of nearly 40{,}000 shells. On the read path, the agent combines structural graph traversal over a Kafka-synchronized knowledge graph with scoped vector search over document content, both grounded in standardized IDTA-template semantics. On the write path, a server-side validator enforces structural template conformance. We evaluate the Qwen3.5 and Qwen3.6 families (2B-397B) to find the minimum model size for these read workflows and the limits of scale for writes. Read-path competence emerges sharply between 9B and 27B. Write-path correctness, however, hits a ceiling that added scale does not lift: structural failures are recoverable through deterministic feedback, whereas controlled-vocabulary gaps are a validator-coverage limit, closed by extending enforcement rather than by scale. These results delineate where larger models help and where deterministic enforcement must reach beyond structure.