Context-Enriched Prompting Strategies for Automated Data Specification Adjustment in Dataspecer Using LLMs, RAG and MCP

This thesis elaborates on the application of LLMs to the automated adjustment of data specifications in response to real-world changes, such as updates of APIs and database schemas. The work is situated in the context of Dataspecer, a tool designed to make data specifications management more intuitive and natural. Research focuses on de- signing and evaluating strategies for LLM input and output to support specification maintenance, with particular attention to prompting techniques such as chaining, role- based prompting, and few-shot examples. The study explores the integration of LLMs into more complex tool environments using the MCP and RAG for the stated purpose. The expected outcomes include a methodology and prototype solution for effective LLM- driven specifications adjustment, alongside an experimental evaluation of the proposed approach.