Agentic GraphRAG untuk Temu Balik Informasi Multi-hop pada Sistem Rekomendasi Buku: Studi Kasus Perpustakaan Publik DKI Jakarta

Fauzia, Mutiara Noor Fauzia (2026) Agentic GraphRAG untuk Temu Balik Informasi Multi-hop pada Sistem Rekomendasi Buku: Studi Kasus Perpustakaan Publik DKI Jakarta. Other thesis, Institut Teknologi Sepuluh Nopember.

Text Buku Tugas Akhir Program Sarjana Sistem Informasi Mutiara Noor Fauzia 5026221045.pdf - Accepted Version Restricted to Repository staff only Download (9MB) | Request a copy

Abstract

Perpustakaan publik menghadapi paradoks efisiensi yakni koleksi yang masif justru menyulitkan temu balik informasi karena sistem katalog OPAC konvensional bertumpu pada pencarian kata kunci yang gagal menangani kueri multi hop berupa pertanyaan yang menuntut penalaran lintas beberapa atribut sekaligus seperti nuansa cerita kemiripan antar buku bahasa dan ketersediaan cabang. Penelitian ini merancang mengimplementasikan dan mengevaluasi arsitektur Agentic GraphRAG untuk temu balik informasi multi hop pada sistem rekomendasi buku dengan studi kasus katalog Perpustakaan Publik Daerah Khusus Ibukota Jakarta. Data katalog diolah melalui saluran data Medallion menjadi Knowledge Graph berkapasitas 16.261 simpul 51.417 relasi dan 4.998 buku yang ditelusuri oleh agen berbasis model bahasa kecil berstatus frozen melalui katalog instrumen Cypher parametrik yang secara mutlak bebas dari risiko halusinasi kueri. Studi ablasi terhadap enam strategi kendali agen pada dua model yakni Llama 3.1 8B dan Qwen 2.5 7B dievaluasi menggunakan kerangka enam metrik hibrida yang mencakup aspek temu balik generasi dan operasional atas seratus kueri ground truth. Hasil pengujian menunjukkan bahwa nilai tambah murni berasal dari kemampuan acting yang menggerakkan rantai temu balik dengan raihan skor Precision@3 antara 0,50 hingga 0,90 apabila disandingkan dengan strategi tanpa acting yang tertahan di kisaran 0,18. Lebih lanjut strategi perencanaan sekali jalan atau Planned pada model Qwen 2.5 memberikan titik keseimbangan akurasi dan efisiensi terbaik dengan capaian Precision@3 di angka 0,903 beserta konsumsi token yang jauh lebih hemat pada tingkatan latensi yang setara dengan metode perulangan reason act.
==============================================================================================================================================
Public libraries face an efficiency paradox: massive collections complicate information retrieval because conventional OPAC catalog systems rely on keyword searches that fail to handle multi-hop queries demanding reasoning across multiple attributes simultaneously, such as story vibes, book similarities, languages, and branch availability. This study designs, implements, and evaluates an Agentic GraphRAG architecture for multi-hop information retrieval in a book recommendation system, using the Jakarta Public Library catalog as a case study. Catalog data is processed through a Medallion data pipeline into a Knowledge Graph containing 16,261 nodes, 51,417 relationships, and 4,998 books, which is navigated by agents based on small, frozen language models via a parametric Cypher tool catalog strictly free from query hallucination risks. An ablation study of six agent control strategies across two models, namely Llama 3.1 8B and Qwen 2.5 7B, is evaluated using a framework of six hybrid metrics covering retrieval, generation, and operational aspects against one hundred ground truth queries. The evaluation results indicate that the pure value addition stems from the acting capability that drives the retrieval chain, achieving Precision@3 scores in the range of 0.50 to 0.90, compared to non-acting strategies that stagnate at around 0.18. Furthermore, the single-shot planning strategy, or Planned, on the Qwen 2.5 model provides the optimal balance of accuracy and efficiency, achieving a Precision@3 of 0.903 along with significantly lower token consumption at a latency level equivalent to the iterative reason-act method.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Agentic GraphRAG, Book Recommender Systems, Knowledge Graph, Large Language Model, Multi-hop Reasoning, Agentic GraphRAG, Graf Pengetahuan, Large Language Model, Multi-hop Reasoning, Sistem Rekomendasi Buku
Subjects:	Q Science > QA Mathematics > QA76.9.I58 Recommender systems (Information filtering)
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Information System > 57201-(S1) Undergraduate Thesis
Depositing User:	Mutiara Noor Fauzia
Date Deposited:	03 Aug 2026 01:02
Last Modified:	03 Aug 2026 01:02
URI:	http://repository.its.ac.id/id/eprint/142190

Actions (login required)

View Item