Maassobirin, Mukhamad Khafid (2026) OPTIMASI TREE TRAVERSAL BERBASIS INTERLEAVING CHAIN-OF-THOUGHT PADA ARSITEKTUR RETRIEVAL-AUGMENTED GENERATION HIERARKIS. Undergraduate thesis, UPN Veteran Jawa Timur.
|
Text (COVER)
ilovepdf_merged (2)_organized.pdf Download (1MB) | Preview |
|
|
Text (BAB I)
SKRIPSI KHAFID INDONESIA BAB 1.pdf Download (356kB) | Preview |
|
![[img]](https://repository.upnjatim.ac.id/style/images/fileicons/text.png) |
Text (BAB II)
SKRIPSI KHAFID INDONESIA BAB 2.pdf Restricted to Repository staff only until 1 July 2029. Download (1MB) | Request a copy |
|
|
Text (BAB III)
SKRIPSI KHAFID INDONESIA BAB 3.pdf Restricted to Repository staff only until 1 July 2029. Download (1MB) | Request a copy |
|
|
Text (BAB IV)
SKRIPSI KHAFID INDONESIA BAB 4.pdf Restricted to Repository staff only until 1 July 2029. Download (1MB) | Request a copy |
|
|
Text (BAB V)
SKRIPSI KHAFID INDONESIA bab 5.pdf Download (243kB) | Preview |
|
|
Text (DAFTAR PUSTAKA)
SKRIPSI KHAFID INDONESIA DAFTAR PUSTAKA.pdf Download (203kB) | Preview |
|
|
Text (LAMPIRAN)
SKRIPSI KHAFID INDONESIA LAMPIRAN.pdf Restricted to Repository staff only until 1 July 2029. Download (543kB) | Request a copy |
Abstract
Hierarchical RAG systems such as RAPTOR and HIRO are limited by traversal mechanisms that rely solely on static semantic similarity against the initial query, making them unable to adapt to evolving reasoning needs during the exploration process. Static similarity-based retrieval is insufficient for tasks requiring deep reasoning over long documents, as demonstrated by the BRIGHT benchmark which revealed significant performance degradation even among top-performing retrieval models. This study proposes an Interleaving Chain-of-Thought (IRCoT)-based tree traversal mechanism within a hierarchical RAG architecture that integrates dynamic reasoning signals into the Depth-First Search algorithm. The system encompasses three main phases: hierarchical tree construction through embedding, clustering, and LLM-based summarization, followed by traversal using a combined scoring function with adaptive dual-threshold pruning interleaved with reasoning generation at each step, and finally an answer generation phase. The key innovation lies in reasoning signals that evolve dynamically throughout traversal, distinguishing this approach from HIRO which relies on static query representations. Comprehensive evaluation was conducted on four benchmark datasets, namely NarrativeQA, QASPER, QuALITY, and TyDi QA, using multilingual-e5-large-instruct for embedding and Qwen2.5-7B-Instruct-AWQ for generation. Results show that IRCoT improves ROUGE-L on NarrativeQA from 0.1144 to 0.1275 and Answer F1 on QASPER from 0.3135 to 0.3288, but decreases accuracy on QuALITY from 54.98% to 54.49% and Token F1 on TyDi QA from 0.3922 to 0.3862, with computational overhead two to four times slower. Empirical analysis identifies that reasoning signals are effective for documents with dispersed information but offer limited advantages for shallow-structured documents or holistic-comprehension-demanding questions.
| Item Type: | Thesis (Undergraduate) | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Contributors: |
|
||||||||||||
| Subjects: | Q Science > Q Science (General) Q Science > QA Mathematics > QA76 Computer software Q Science > QA Mathematics > QA76.6 Computer Programming Q Science > QA Mathematics > QA76.76.E95 Expert Systems Q Science > QA Mathematics > QA76.87 Neural computers |
||||||||||||
| Divisions: | Faculty of Computer Science > Departemen of Data Science | ||||||||||||
| Depositing User: | Mukhamad Khafid Maassobirin | ||||||||||||
| Date Deposited: | 01 Jul 2026 04:38 | ||||||||||||
| Last Modified: | 01 Jul 2026 04:58 | ||||||||||||
| URI: | https://repository.upnjatim.ac.id/id/eprint/54348 |
Actions (login required)
 |
View Item |