Laajojen kielimallien käyttö kontekstilla avustettuna teollisuusjärjestelmissä

Large Language Models (LLMs) enable natural language interaction with software systems, but their application in industrial environments remains challenging due to strict security requirements, legacy architectures, and domain-specific data schemas. Focusing on the integration of an LLM and a Manufacturing Execution System (MES), the LLM integration requires that the LLM has the ability to access and combine operational data across multiple systems and database tables dynamically, where traditionally premade interfaces allow the user to execute specific queries defined in the API. This thesis investigates how to enable context-aware LLM assistance in industrial systems by integrating an LLM with an MES using the Model Context Protocol (MCP). Utilizing the Design Science Research Methodology, a proof-of-concept artifact was designed and implemented where an MCP client connects user prompts to an LLM and an MCP server exposes read-only database tools for schema discovery and SQL execution. The prototype aims to support data retrieval from multiple sources, here demonstrated with multiple database tables that are queried with joins or by executing multiple queries, and to parse natural language time expressions into SQL queries with correct timestamp ranges, while enforcing guardrails that prevent altering or destructive operations on existing data. Artifact capabilities are demonstrated through realistic MES query scenarios and evaluated using automated and manual end-to-end tests with both a Azure cloud hosted model (GPT-4o) and a locally Ollama deployed (GPT-OSS:20B). The results show that MCP based LLM integration to industrial systems can enable context-aware data access utilizing tools without modifying the MES architecture, and that model choice and deployment strongly affect latency and reliability of LLM queries and output. The evaluation highlights time expression parsing and ambiguity handling as the main remaining limitations for reliable industrial use of LLMs.