Pengaruh Pemilihan Vector Database terhadap Skalabilitas dan Kecepatan Respons Berbasis Retrieval-Augmented Generation

Farisy, Muhammad Mirza (2026) Pengaruh Pemilihan Vector Database terhadap Skalabilitas dan Kecepatan Respons Berbasis Retrieval-Augmented Generation. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5024221053-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (1MB) | Request a copy

Abstract

Sistem chatbot helpdesk berbasis Retrieval-Augmented Generation (RAG) membutuhkan proses temu kembali dokumen yang cepat, skalabel, dan relevan agar mampu menghasilkan jawaban yang responsif serta dapat ditelusuri sumbernya. Salah satu komponen penting dalam proses tersebut adalah vector database, karena komponen ini menyimpan representasi vektor dokumen dan menjalankan pencarian semantik terhadap pertanyaan pengguna. Penelitian ini bertujuan menganalisis pengaruh pemilihan vector database terhadap kecepatan respons, skalabilitas, dan kualitas temu kembali pada sistem RAG untuk domain helpdesk. Penelitian dilakukan dengan membandingkan enam vector database, yaitu ChromaDB, Qdrant, LanceDB, PostgreSQL dengan pgvector, SQLite, dan Pinecone, pada rancangan sistem RAG yang sama. Pengujian menggunakan model embedding qwen3-embedding:8b dan model bahasa gemma4:e4b untuk pembangkitan jawaban. Evaluasi mencakup waktu temu kembali, total waktu respons, sensitivitas nilai Top-K, indikasi skalabilitas terhadap ukuran corpus hingga 15.631 chunk, skalabilitas terhadap jumlah pengguna konkuren, serta kualitas temu kembali dan jawaban menggunakan kerangka DeepEval. Hasil pengujian menunjukkan bahwa ChromaDB dan Qdrant secara konsisten memberikan temu kembali tercepat, paling stabil terhadap pertumbuhan corpus, serta throughput tertinggi tanpa kegagalan pada pengujian pengguna konkuren. Sebaliknya, SQLite menurun tajam di bawah beban paralel, sedangkan Pinecone paling lambat, dengan pengukuran yang mencakup latensi jaringan sebagai layanan cloud. Skor kualitas DeepEval antar database relatif berdekatan dan tidak menunjukkan perbedaan kualitas yang
dapat diandalkan antar database pada subset kueri yang dievaluasi. Temuan ini menunjukkan bahwa pemilihan vector database terutama memengaruhi kecepatan dan skalabilitas, sedangkan perbedaan kualitas antar database tidak teramati pada konfigurasi model embedding yang ditahan tetap.
=====================================================================================================================================
A Retrieval-Augmented Generation (RAG)-based helpdesk chatbot requires fast, scalable, and relevant document retrieval in order to generate responsive answers with traceable sources. One important component in this process is the vector database, which stores document vector
representations and performs semantic search for user queries. This research analyzes the effect of vector database selection on response speed, scalability, and retrieval quality in a RAG-based helpdesk system. The study compares six vector databases, namely ChromaDB, Qdrant, LanceDB, Post- greSQL with pgvector, SQLite, and Pinecone, under the same RAG system design. The evaluation uses the qwen3-embedding:8b embedding model with the gemma4:e4b language model for answer generation. Experiments cover retrieval time, total response time, Top-K sensitivity, scalability against a corpus of up to 15,631 chunks, scalability against the number of concurrent users, and retrieval and answer quality using the DeepEval framework. The results show that ChromaDB and Qdrant consistently provide the fastest retrieval, the highest stability against corpus growth, and the highest throughput with no failures under concurrent-user testing. In contrast, SQLite degrades sharply under parallel load, while Pinecone is the slowest, with measurements that include network latency as a managed cloud service. DeepEval quality scores across databases are relatively close and show no reliable quality differences between databases on the evaluated query subset. These findings indicate that vector database selection mainly affects speed and scalability, while no database-level quality differences were observed under the fixed embedding configuration.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Vector Database, Retrieval-Augmented Generation, Chatbot Helpdesk, Skalabilitas, ChromaDB, Qdrant, Vector Database, Retrieval-Augmented Generation, Helpdesk Chatbot, Scalability, ChromaDB, Qdrant.
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence Z Bibliography. Library Science. Information Resources > ZA Information resources > Z699.5 Information storage and retrieval systems Z Bibliography. Library Science. Information Resources > ZA Information resources > ZA4450 Databases
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Computer Engineering > 90243-(S1) Undergraduate Thesis
Depositing User:	Muhammad Mirza Farisy
Date Deposited:	22 Jul 2026 08:31
Last Modified:	22 Jul 2026 08:31
URI:	http://repository.its.ac.id/id/eprint/136327

Actions (login required)

View Item