Yulestiono, Achmad Yusuf (2026) Implementation of Structure-Aware Hybrid Chunking in Retrieval-Augmented Generation (RAG) for Small Language Model (SLM) in Surabaya City Regional Tax Regulation Documents. Undergraduate thesis, upn jatim.
|
Text (Cover)
22081010180-cover.pdf Download (1MB) | Preview |
|
|
Text (Bab 1)
22081010180-bab1.pdf Download (29kB) | Preview |
|
![[img]](https://repository.upnjatim.ac.id/style/images/fileicons/text.png) |
Text (Bab 2)
22081010180-bab2.pdf Restricted to Repository staff only until 29 June 2029. Download (466kB) |
|
|
Text (Bab 3)
22081010180-bab3.pdf Restricted to Repository staff only until 29 June 2029. Download (1MB) |
|
|
Text (Bab 4)
22081010180-bab4.pdf Restricted to Repository staff only until 29 June 2029. Download (1MB) |
|
|
Text (Bab 5)
22081010180-bab5.pdf Download (20kB) | Preview |
|
|
Text (Daftar Pustaka)
22081010180-daftarpustaka.pdf Download (89kB) | Preview |
|
|
Text (Lampiran)
22081010180-lampiran.pdf Restricted to Repository staff only until 29 June 2029. Download (147kB) |
Abstract
Local tax regulatory documents have a complex hierarchy that makes it difficult to find keyword-based information. This study implements Structure-Aware Hybrid Chunking (SAHC) on the Retrieval-Augmented Generation (RAG) architecture to break down documents based on the boundaries of legal structures (Chapters, Articles, Verses) and semantics, using the Small Language Model Phi-3.5. The performance of SAHC was compared to fixed-size and semantic chunking through three test scenarios. The results showed that SAHC with hybrid Reciprocal Rank Fusion (RRF) achieved the best retrieval performance with Precision@5 (0.3638), Recall@5 (0.7632), and F1@5 (0.4596). The high quality of evidence search is able to support the accuracy of answers, although the stability of text generation requires strict evidence screening . In addition, corpus updates have proven to be more efficient using incremental dense indexing. In conclusion, SAHC proves to be an optimal segmentation approach for hierarchical documents because it is able to maintain structural integrity and ensure that answers can be traced back to their original normative sources.
| Item Type: | Thesis (Undergraduate) | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Contributors: |
|
||||||||||||
| Subjects: | Q Science > QA Mathematics > QA76.87 Neural computers | ||||||||||||
| Depositing User: | Achmad Yusuf Yulestiono | ||||||||||||
| Date Deposited: | 29 Jun 2026 06:57 | ||||||||||||
| Last Modified: | 29 Jun 2026 07:51 | ||||||||||||
| URI: | https://repository.upnjatim.ac.id/id/eprint/54287 |
Actions (login required)
 |
View Item |