LLM ALIGNMENT WITH PROMPT ENGINEERING AND RETRIEVAL AUGMENTED GENERATION (CASE STUDY: CONVERSATIONAL AGENT FOR THE ”MIST” DIARY SYSTEM)

• Main Author: Attarizal Rezyarifin, Zhillan
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/85486
• Institution: Institut Teknologi Bandung

Aligning LLM (Large Language Models) for specific needs and contexts can significantly enhance the quality of user experience and the growth of LLM-based applications. Unfortunately, LLM alignment methods are still new, limited, and require experimental development. Prompt engineering and retrieval-augmented generation are two promising LLM alignment methods that can be utilized by developers who do not have extensive computational resources. This thesis aligns an LLM to meet the needs of the ”MIST” diary system conversational agent case study. A full-stack web application was chosen as the instrument for implementing the LLM alignment. The implementation includes deployment on hardware, development of inference services, RAG (retrieval- augmented generation), prompt engineering, memory, interface, and the creation of a large number of prompt contexts. System evaluation shows that all implemented functionalities were successfully executed, and the context was successfully injected to control the behavior and interaction style of MIST. Unfortunately, qualitative evaluation with users indicated that the output quality, intelligence level, and system latency were still unsatisfactory. Further development is needed regarding LLM inference optimization and fast, iterative methods for creating and testing prompt templates.