Pengembangan Sistem Rekomendasi Mobil Toyota Berbasis Hybrid Retrieval-Augmented Generation Menggunakan TiDB Vector Search dan Google Gemini

Ramadani, M.Riski (2026) Pengembangan Sistem Rekomendasi Mobil Toyota Berbasis Hybrid Retrieval-Augmented Generation Menggunakan TiDB Vector Search dan Google Gemini. Skripsi thesis, Universitas Islam Negeri Imam Bonjol Padang.

	Text (cover) M.Riski Ramadani-2217020018-cover- daftar isi.pdf - Published Version Download (1MB)
	Text (Bab-1) bab-1.pdf - Published Version Download (2MB)
	Text (Bab-3) bab-3.pdf - Published Version Download (4MB)
	Text (Bab-5-dan-Daftar-Pustaka) bab-5-dafpus.pdf - Published Version Download (2MB)
	Text (fulltext) M.Riski Ramadani-2217020018-Full Text Skripsi.pdf - Published Version Restricted to Repository staff only Download (13MB)

Abstract

Toyota's complex and continually changing range of models and specifications makes it difficult for prospective customers to obtain accurate, needs-appropriate information. The digital assistant on Auto2000 Rantauprapat's website previously relied on static prompt-based knowledge injection, embedding knowledge directly into the prompt without a dynamic retrieval mechanism, making it prone to hallucination, particularly for frequently changing OTR pricing and vehicle specification data. This study develops and integrates a Hybrid Retrieval-Augmented Generation (RAG) architecture into the website's recommendation system and chatbot to improve response accuracy and eliminate hallucination. The architecture begins with a Self-Querying Retriever, powered by the Gemini model, that decomposes natural-language queries into structured parameters. These parameters are executed through two parallel search paths on TiDB Serverless: Dense Vector Search for semantic similarity and Sparse SQL Search for structured filtering, before being merged, scored via BM25, and re-ranked through Reciprocal Rank Fusion (RRF). Development includes data ingestion from brochures and price lists, entity-based chunking to preserve each vehicle unit's context, embedding generation with the BGE-M3 model, integration of Google Gemini for conversational responses, and a Flask-based admin portal with automatic brochure extraction via Gemini Multimodal OCR to accelerate catalog management. The system was evaluated through three complementary methods: retrieval and answer accuracy using the DeepEval framework across 25 test scenarios; Black Box testing across 31 functional scenarios (18 consumer-module, 13 admin-module); and usability evaluation via the System Usability Scale (SUS) with 15 respondents. Results show an average DeepEval score of 86.79%, a perfect Faithfulness score (1.00) across all scenarios, a 100% Black Box pass rate, and an SUS score of 85.67, classified as Excellent (Grade A). These findings confirm that the system mitigates hallucination risk, performs according to its design across consumer and admin features, and is well received by users, demonstrating its feasibility as a product consultation tool at Auto2000 Rantauprapat.

Actions (login required)

View Item