Retrieval-Augmented Generation in Corporate Environments: Development and Testing of an HR Conversational Agent

This thesis describes the design, development and testing of a Proof of Concept of an HR chatbot, carried out in 2025 at Amadori, an Italian agri-food company, together with its People & Organization function. At Amadori, employees often asked the HR department questions whose answers were already contained in company documents. The chatbot, available in Google Chat, follows the Retrieval-Augmented Generation (RAG) approach: it retrieves passages from the HR documents of the company and uses a large language model (Gemini) to generate answers based on them. It was built on Google Cloud with Dialogflow CX and Vertex AI Search. Two prototypes were developed and compared: one based on documents converted to HTML and split into chunks, the other on the original PDF files. Since the access control lists planned for confidential documents did not work as expected, the answers that depend on the collective labour agreement of the employee were handled through prompt engineering. The project was managed with an adaptive project management life cycle, using a Scope Bank to handle the changes of scope. The HTML prototype was tested with 151 real requests sent by the HR team, which were then replayed automatically on the PDF prototype. The HTML chatbot answered 78.1% of the questions and the PDF chatbot 72.7%; most of the unanswered questions were outside the scope of the documents. Since the results were similar and PDF documents are easier to maintain, the PDF chatbot was chosen. The POC was approved by the business. The thesis also discusses the lessons learned and the limits of agile approaches in more complex organisational contexts.