Implementasi Semantic Retrieval Berbasis Embedding dan LLM pada Aplikasi Web Keuangan Pribadi

Abstract

This study developed a web-based personal finance application integrated with semantic retrieval using embedding and Large Language Models to address limitations in conventional keyword-based search. The system was developed using the Extreme Programming approach to ensure iterative development, rapid feedback, and adaptability to user requirements. The application was designed using a client–server architecture and implemented to support transaction recording, financial management, and contextual data retrieval. Financial transaction data were preprocessed and converted into vector embeddings to capture semantic meaning, enabling more flexible and accurate search using natural language queries. The system utilized a vector database with pgvector extension to store embeddings and perform similarity search. Evaluation was conducted using the RAGAs framework with 100 financial transaction data and 20 query scenarios. The results showed that the system achieved a context recall of 0.81, context precision of 0.70, faithfulness of 0.83, and answer relevancy of 0.84. These results indicate that the system was able to retrieve relevant contextual information and generate consistent and accurate responses. Furthermore, the implementation demonstrated improved effectiveness in organizing, managing, and retrieving financial data compared to traditional methods. Overall, the proposed system successfully provided a more efficient and context-aware solution for personal financial management and information retrieval.