KMITL

Permanent URI for this communityhttps://dspace.kmitl.ac.th/handle/kmitl/1

Browse

Settings

Subject: search.filters.subject.Retrieval-Augmented Generation

Search Results

Now showing 1 - 3 of 3
  • Loading...
    Thumbnail Image
    Some of the metrics are blocked by your 
    consent settings
    Item type:Publication,
    MANJU: A Multi-Agent Framework for Natural Just-in-Time Understanding in Thai
    (2026-01-01)
    Koseeyaumporn, Sawit
    ;
    Saiprom, Siratee
    ;
    Promfiy, Watin
    ;
    Tarnpradab, Sansiri
    This paper presents MANJU (Multi-Agent AI for Natural Just-in-Time Understanding), an automated platform for Thai-language customer service delivered through just-in-time voice interaction. MANJU transcribes a customer's spoken query, applies multi-agent reasoning with retrieval-augmented generation (RAG) to determine the optimal response strategy, and synthesises a reply in natural Thai, delivering the first audio segment within approximately 1.8 s via sentence-level streaming. A visual drag-and-drop work flow editor enables non-technical operators to compose and modify conversation logic without programming.Evaluated on 50 standard Thai customer-service queries and 23 adversarial edge cases, MANJU achieves 100% intent routing accuracy, 62% RAG-grounded responses (GR), 78% edge-case acceptability with 100% safety, and Mean Opinion Score (MOS) = 4.33 for VoiceDesign TTS. End-to-end median latency is 6.07 s; the primary bottleneck is Qwen3-TTS at Real-Time Factor (RTF) = 1.54, partially offset by sentence-level streaming.
  • Loading...
    Thumbnail Image
    Some of the metrics are blocked by your 
    consent settings
    Item type:Publication,
    Quantitative Assessment of Retrieval Strategies in RAG Architectures: A Comparative Study Across Multiple Knowledge Domains Using Standardized Performance Metrics
    (2025-01-01)
    Aromsuk, Tinnarat
    ;
    Nootyaskool, Supakit
    ;
    Netisopakul, Ponrudee
    This study conducts a systematic quantitative assessment of four distinct retrieval methodologies within Retrieval-Augmented Generation (RAG) frameworks: baseline implementation, hybrid dual-paradigm approach, hierarchical parent-child structure, and contextual compression. Through rigorous experimental evaluation spanning six distinct knowledge domains, we employ established metrics including ROUGE [1], BLEU [2], and computational timing measurements to characterize performance profiles. Our findings reveal that sophisticated retrieval approaches deliver substantial computational efficiency gains (4-5 × acceleration) alongside varied performance patterns across quality assessment dimensions. The hierarchical parent-child methodology demonstrates superior BLEU performance (0.1046 mean score) coupled with optimal retrieval speeds (0.0124 s), whereas hybrid approaches excel in ROUGE metrics (0.0317 mean score). Domain-specific analysis indicates pronounced performance disparities: medical/health domains (COVID19 pandemic) achieve highest aggregate scores (0.1198 mean ROUGE), while specialized technical and legal domains present distinct retrieval complexities. This research establishes empirical foundations for evidence-based retrieval method selection, identifying clear efficiency-quality relationships and domain-dependent optimization strategies for production RAG deployments.
  • Loading...
    Thumbnail Image
    Some of the metrics are blocked by your 
    consent settings
    Item type:Publication,
    Thai Question-Answering System Using Similarity Search and LLM
    (2024-07-01)
    Jearanaitanakij, Kietikul
    ;
    Srithongdee, Chananchida
    ;
    Ketkham, Sirinoot
    ;
    Ardsana, Onwanya
    ;
    Kullawan, Tiwat
    A question-answering (QA) system is essential to an organization where numerous QA pairs respond to customer queries. Choosing the right pair corresponding to the query is a complex task. Although the QA system from a commercial product like ChatGPT provides an excellent solution, it is costly, and the fine-tuned Large Language Model (LLM) cannot be downloaded for private use at the local site. In addition, the cost of using such LLM may significantly increase when the number of users grows. We propose a Thai QA system that can swiftly respond and correctly match the user query to the reference answer in the QA dataset. The proposed system encodes both QA pairs and a query into individual embeddings and finds a couple of QA pairs that are most related to the query by using the fast similarity search called Faiss (Facebook AI Similarity Search.) Afterward, the relevant QA pairs and the query are fed to the fine-tuned LLM (WangchanBERTa-pretraining multilingual transformer-based) to choose the single best match QA pair. The fine-tuned WangchanBERTa can retrieve the correct answer and respond to the query naturally. The experiment conducted on the Thai Wiki QA dataset indicates the superior ROUGE values, precision, recall, F1-score, and runtime of the proposed system against other strategies.