Year: 2026
Type: dissertation
Abstract: Customer Relationship Management (CRM) solutions increasingly integrate AI chatbots.The requirements for information retrieval are complex, making the selection of the optimal database type critical.This thesis analyzes how different database types compare in their retrieval capabilities within a CRM context.To solve the initiating problem we constructed two research questions.The first RQ examines how CRM system requirements shape the implementation of an chatbot.Our second RQ investigates the performance differences between alternative database types based on the system requirements.In our first phase we conducted interviews with stakeholders to gain insight into use cases and system requirements.These insights were utilized in Phase 2 that was an experimental phase over four iterations in which each completed iteration was evaluated and new exploratory ideas were carried on.The results demonstrated that while graph databases generally delivered strong overall performance and high accuracy by utilizing graph traversal, vector databases outperformed them in certain scenarios specifically on retrieval from unstructured data and semantic prompts.These findings suggest that a combined approach with a graph and vector database is a suitable choice for this context.The evaluation demonstrates that minimizing tool calls is the most critical factor for designing an efficient retriever.Since the context must be repeatedly transmitted back and forth with each tool call, excessive calls significantly degrade overall performance and latency.
Authors: Fredlund, Martin, Herstedt, Magnus
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7155025195