Rancang Bangun Sistem Analitik Data Terstruktur Berbasis Orkestrasi Sekuensial Agen Small Language Model Lokal Terkuantisasi

Syahputra, Yoga Firman (2026) Rancang Bangun Sistem Analitik Data Terstruktur Berbasis Orkestrasi Sekuensial Agen Small Language Model Lokal Terkuantisasi. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025221212-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (4MB) | Request a copy

Abstract

Pertumbuhan volume data terstruktur mendorong kebutuhan akan sistem analitik yang dapat diakses oleh pengguna non-teknis melalui text-to-SQL. Namun, pemanfaatannya yang bertumpu pada Large Language Model (LLM) berbasis awan memicu kendala biaya tinggi, ketergantungan konektivitas, dan risiko privasi data. Sebagai alternatif, Small Language Model (SLM) lokal terkuantisasi mampu menjaga data tetap aman di perangkat pengguna. Penelitian ini merancang dan membangun sistem analitik data terstruktur end-to-end berbasis orkestrasi sekuensial agen menggunakan satu SLM lokal terkuantisasi pada laptop konsumer (RAM 8 GB, GPU GTX 1650 4 GB) dengan memadukan DuckDB, ChromaDB untuk retrieval-augmented generation, scaffolding deterministik (rule-based resolver dan semantic guard), serta fine tuning QLoRA untuk agen penyusun narasi. Pengujian dilakukan pada domain energi dan finansial melalui delapan skenario eksperimen yang dievaluasi menggunakan execution accuracy, uji McNemar, dan bootstrap confidence interval. Hasil penelitian menunjukkan sistem berhasil berjalan penuh secara lokal tanpa kehabisan memori GPU, mencapai execution success rate 96,15% meskipun full accuracy pada domain energi hanya 30,77% yang mengindikasikan persoalan utama pada ketepatan semantik, bukan validitas kueri. Selain itu, teknik prompting terbukti berkontribusi jauh lebih besar daripada scaffolding deterministik, di mana strategi few-shot statis mencapai full accuracy 54,81% melampaui sistem berkonfigurasi lengkap, sementara eksekusi GPU terbukti 2,59 kali lebih cepat dibanding CPU dengan inferensi model menyerap 96,8% total waktu eksekusi. Hal ini membuktikan kelayakan SLM lokal terkuantisasi sebagai fondasi analitik data pada perangkat terbatas, meskipun kinerjanya sangat bergantung pada rancangan prompt dan scaffolding, bukan pada kemampuan model semata.
==================================================================================================================================
The growth of structured data volume drives the need for analytical systems accessible to non-technical users through text-to-SQL. However, relying on cloud-based Large Language Models (LLMs) poses challenges such as high costs, connectivity dependencies, and data privacy risks. As an alternative, local quantized Small Language Models (SLMs) keep data secure on the user's device. This study designs and builds an end-to-end structured data analytics system based on sequential agent orchestration using a single local quantized SLM on a consumer laptop (8 GB RAM, GPU GTX 1650 4 GB), integrating DuckDB, ChromaDB for retrieval-augmented generation, deterministic scaffolding (a rule-based resolver and a semantic guard), and QLoRA fine-tuning for the narrative agent. Testing was conducted across the energy and finance domains through eight experimental scenarios evaluated using execution accuracy, exact McNemar's test, and bootstrap confidence intervals. Results show that the system successfully runs fully locally without GPU memory overflow, achieving a 96.15% execution success rate although full accuracy in the energy domain was only 30.77%, indicating that the primary issue lies in semantic precision rather than query validity. Furthermore, prompting techniques proved to contribute significantly more than deterministic scaffolding, where a static few-shot strategy reached 54.81% full accuracy, outperforming the fully configured system, while GPU execution was 2.59 times faster than CPU, with model inference consuming 96.8% of total execution time. This demonstrates the viability of local quantized SLMs as a foundation for data analytics on resource-constrained devices, although their performance depends heavily on prompt design and scaffolding rather than model capability alone.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Small Language Model, analitik data terstruktur, text-to-SQL, orkestrasi agen, kuantisasi, komputasi lokal, Small Language Model, structured data analytics, text-to-SQL, agent orchestration, quantization, local computing.
Subjects:	T Technology > T Technology (General) > T58.62 Decision support systems
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Yoga Firman Syahputra
Date Deposited:	03 Aug 2026 09:39
Last Modified:	03 Aug 2026 09:39
URI:	http://repository.its.ac.id/id/eprint/142469

Actions (login required)

View Item