Development of a conversational tool based on Retrieval-Augmented Generation for leveraging internal data within the company Connectic

Details

Supervisors

Schaus, Pierre

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] en sciences informatiques, à finalité spécialisée

Abstract

enCompanies increasingly store large amounts of operational data in enterprise systems such as ERPs. However, this information is sparsely distributed across several sources, which makes it difficult for employees to retrieve quickly through traditional interfaces. This thesis investigates the design, implementation and evaluation of a Retrieval-Augmented Generation-based conversational assistant for Connectic, a Belgian-French digital integrator specialised in IoT solutions and Odoo integration. The objective of the proposed system is to allow Connectic employees to ask natural language questions over internal Odoo data. The assistant relies on an architecture in which data is extracted from an Odoo PostgreSQL dump, transformed into natural-language semantic documents, indexed in OpenAI vector stores, retrieved at query time and injected into the prompt of a Large Language Model. To improve the management of heterogeneous data sources, the documents are separated into multiple vector stores and a retrieval routing layer is introduced to predict the most relevant stores for each query. The system also includes mechanisms for conversation management, source citation, incremental data maintenance and a proposed access control strategy based on Odoo permissions. The assistant is evaluated on a dataset of 47 handcrafted use cases designed with Connectic. The evaluation is conducted layer by layer, by separately measuring routing, retrieval and generation, before comparing the complete pipeline with an other system achieving the same goal. The routing and generation components of the system shows strong results while retrieval remains the main bottleneck of the system. This limitation affects end-to-end performance. The proposed RAG assistant remains more controllable and significantly more cost-efficient, while providing a clear architecture in which each component can be analysed and improved independently. This work therefore demonstrates the feasibility of building a RAG-based conversational assistant over internal Odoo data, while highlighting the importance of robust retrieval, access control and evaluation methodology for enterprise RAG systems.