A Human-Centred Architecture for Large Language Models-Cognitive Assistants in Manufacturing within Quality Management Systems

Preprint A Human-Centred Arc hitecture for Large Language Models-Cogni tive Assistants in Man ufacturing within Quality Managem ent Systems Marcos Galdino a *, Johanna Grahl a , Tobias Hamann a , Anas Abdelrazeq a , Ingrid Isenhardt a a IQS Intelligence in Quality Sens ing, Laborat ory for Machine T ools and Production Engine ering (WZL) of R WTH Aach en University, RWTH Aachen U niversity, Aachen, G ermany * Correspondin g author. Tel.: +49- 178 - 67 -32321; E-mail address: marcos. galdino@wzl-iqs.rwth-aac hen.de Abstract Large Language Models-Cognitive Assistants (LLM-CAs) can enhance Quality Management Systems (QMS) in manufacturing, fostering continuous p rocess improvement and knowledge management. However, there is no hu man-centred software architecture focused on QMS that enables the in tegration of LLM-CAs in to manufacturing in the current literature. This study addresses this g ap by designing a component-b ased architecture considering requireme nt analysis and software development process. Validation was conducted via iterative expert focus groups. The proposed architecture ensures flexibility, scalability, mo dularity, and wor k augmentation within QMS. Mo reover, it paves the way for its operationalization with industrial partners, showcasing its potential for advancing manufacturing processes. Keywords: "Large language models-cogniti ve assistants; qua lity manageme nt systems; manufa cturing; softwar e architecture" 1. Introduction Large Language Model s-Cognitive Assistants (LL M-CAs) can contribute to improving operational performan ce in manufacturing within Quality Managemen t Systems (QMS) [1]. In this con text, LLM-CAs must also fulfill Q MS requirements, on ce being part of it. LLM -CAs are artificial inte lligence -driven systems that process and generate h uman -like text often used to support decision-m aking, automation, and comm unication. They bridge humans and technology, augmenting cognitive tasks and operational pro cesses, e.g., knowled ge sharing [2,3]. However, their implementation in manufacturing still f aces challen ges , e.g., hallucination [4] and lack of data privac y and security [5] . QMS, as per ISO 9001:2015 , provide a structured fram ework to ensure consistent product quality, regulatory compliance, and continu ous im provement [6]. One of the pivotal aspects of QMS is to ensure that kn owledge is available to achieve conformity o f products and services [7] . Although many authors address the development of LLM- CAs, none provides a holistic software arch itecture solution focus ed on th eir implem entation for manufacturing within QMS. Moreov er, the current solutions do not combine fine tuning and retrieval augmented g eneration (RAG), expertise- based h ierarchical upd ate of kn owledge bases , comp liance, fact, and jailbreak checks . Such requirements have been reported in the mentioned domain through different studies [1] . Besides, a central po int in this study is that QM S require a structured document management a p proach, once documentation is sub jected to au dits. Th e implemen tation of LLM -CAs within QMS assum es, therefo re, th e existence of an extensive, versioned data corpora, e.g ., work instructions, best practices, and machine manuals. Combining LLM- CA s and QMS could gu arantee a smooth integration of LLM-CAs in manufacturing as it would ben efit from well stablished industrial standards. Therefore, we believ e t hat a succ essful implementation of LLM -CAs in manufacturing builds upon the principles of QM S . On that basis, this study aims to pr esent an LLM -CA software architecture for QMS that enables efficient and reliab le hum an-centred knowledge management and continuous improvement. To achieve that, the f ollowing research qu estion will be answered: How can a system architectu re for hu man- centred LLM- CA s in manufa cturing within QMS be designed ? T o answer this question a requirem ent analysis for LLM- CAs in manufacturing within QMS carried out by Galdino et al. [1] is co nsidered followed by software dev elopment process and validation th rough focus groups. This pap er is organised as follows: Chap ter 2 provides the theoretical backg round reg arding L LM-CAs, QMS, an d cu rrent solutions for LLM -CAs. Chapter 3 outlines the m ethods. Chapter 4 presents the results, prov iding detailed dep iction and explanation of the developed architectu re. Finally, Chapter 5 concludes the paper contextualizing the architecture within 2 QMS, sum marizing the key findings and providing an outlook for future research. 2. Theoretical Backg round 2.1. Large Lang uage Model s-Cognitive Assistants Large language models ( LLM s) transformed natural language processing (NLP) by developing advanced architectures and vast datasets to generate human- like communication [8]. This developmen t was enabled by the introduction of the machine-learning T ransformer architecture by Vaswan i et al. [9]. The Transformer ar chitecture u tilizes an attention mechanism that allo ws for parallel processing instead of sequen tial computation [ 9]. For instance , Fan et al . [ 10] explored th e potential of LLM agents for industrial robotics. In robotic task p lanning, they achieved a success rate of 82%, u sing a GPT-4 agent . Kernan Freire et al. [11] proposed an LLM -CA in a factory setting aiming at answer ing q ueries an d sh aring of kno wledge. In this setting, GPT-4 dem onstrated superior performan ce, closely followed by open -source LLMs, e.g., StableBelug a2 and Mixtral 8x7 B. LLM -CAs p resent a p otential d isruptive impact in manufacturing . On the one hand, their d eployment still f aces various technical challen ges, e.g., reasoning errors [3]. On the other hand, human and organ isational ch allenges, e.g., artificial intelligence avoidance and lack of skilled wo rkforce mu st also be considered [12]. On that basis, the deployment o f LLM -CAs depend on their effective interp lay with the subsystems “humans, techn ology, and organisation” [1 ]. 2.2. Quality Man agement Systems (ISO 9001 :2015) QMS, as defined by ISO 9001:2015, establish a structured approach to ensurin g consistent quality and organizational excellence [13]. The cu rrent version of ISO 9001:2015 emphasizes risk -based th inking, leadership commitment, stakeholder engagement, process managem ent, and continuous improvem ent [14]. According to ISO 9 001:2015 [7], a QMS defines objectiv es, manages resource s and proce sses, optimizes decision-m aking, a n d iden tifies action s to address in tended and unintended consequences. Th is ap proach guides organizations in developing, implementing, and improving their QMS to align with strategic goals and custom er requiremen ts [13]. Empirical studies confirm the b enefits of QMS adoption. For in stance, So lomon et al. [15] found that QMS improved efficiency and sustainab ility in the electricity sector by fostering quality cu lture and process optimization. More recently, Mu stroph and Rinderle- Ma [16] con ceptualized a software as a service-based QMS aligned with the EU Artificial Intelligence Act (EU AI Act) to monitor artificial intelligence system design, q uality, and risk manag ement. Despite its broad applicability across industries, QMS implementation presents challenges , e.g., innovation and workforce engagem ent, cyb ersecurity an d data pro tection , and resistance to ch ange and leadership issues [1 7]. Ho wever, as manufacturing advances, the adoption of QMS remains crucial for sustain ing co mpetitiveness and ensuring quality -driven technolog ical integration [18]. 2.3. Current solution s for LLM-CA s The cu rrent liter ature on LLM -CAs presents var ious solutions and architectures for their im plementation. Kernan Freir e [2] implemented an LLM-CA in a factory setting for knowledge sharing. His LLM -CA dynamically updates the k nowledge base with operator kno wledge by means of a knowledge graph and dialectic interaction. His results highlight an overall positive em ployee per ception of the LLM- CA in a detergent facto ry. Emp loyees reported various benefits, e.g., improved kn owledge sharing and problem - solving. Ho wever, his tech nical ap proach does no t comprise mechanisms for safeguarding knowledge quality, approving procedures, com pliance check, an d guaranteeing that the knowledg e base is always up to date. Bucaioni et al. [19] pro posed a ref erence architecture for the integration of LLM s in to software systems. First, they identified arc hitectural concerns regarding LLMs, e. g., data handling, performance , scalability, amon g oth ers. Second, based on the concerns, they developed a modular architecture , which aims at pro viding a general ap proach for LLM integration. Nev ertheless, th ere is no elucidation of user feedback m echanisms. In addition, their architecture does not comprise access control th at regulates user interaction within the architecture. Their solution co uld be transferred to QMS . However, missing specific requirements, e.g., documen tation managemen t, presents potential furth er development. The previous two examples demonstrate that research on LLM integr ation has ad vanced in the last years. However, th e synergy b etween LLMs as an em erging techno logy and the specific requir ements of QMS is still missing. 3. Methods The p roposed ar chitecture has been d eveloped using software d evelopmen t process, utilizing Domain- Driven Design (DDD) . Th e result is a micro service-b ased mu lti-agent system (MAS). Software dev elopment is a concep tual, iterative proce ss. It typically consists of the phases: requiremen t analysis, design, implementation and testing [20,2 1]. Our work in this paper covers the r equirement analysis and design phase. Requirements an alysis marks the beginning o f the developm ent process an d involv es th e exam ination of the problem [21] . Requir ements extracted in this phase form the basis for the system to be built. T hey are d ivided into functional (FR) and non-function al (NFR) requ irements . FRs define the functionality of the sy stem, e.g., the system shall inco rporate human feedback into its knowledge base, whereas NFRs define the system qualities and p roperties, e.g., the system should be scalable and adaptab le [22] . Table 1 presen ts the fundamental requirements of ou r system, wh ich are the basis of our cognitive assistant system d esign. These requirements originate 3 from an earlier work [ 1] and represen t the techn ically operationaliza ble requirements. On that basis, more detailed requirements were defined, which were then div ided into FR and NFR and . Table 1: Requirements operationalized into the LL M-CA softw are architecture (Adapted from Ga ldino et al. [1]) ID Requirement 1 Enable artificial intelligence trustwort hiness 2 Improve resista nce to adversari al input 3 Enhance adapta b ility and scala bility 4 Ensure reliable mo del performance 5 Guarantee effec tive memory usa g e 6 Ensure accurate language output 7 Ensure robus t language understandin g 8 Include human- in -the-loo p 9 Design user-centri c cognitive assista nts 10 Enable effecti ve communication-driv en decision-making 11 Develop a skille d workforc e 12 Adapt to fact ory environments 13 Integrate industr y- sp ecific knowled ge 14 Ensure complian ce integrity In the design phase, a solution to the problem , i.e., identified research gap, defined by the requiremen ts is cr eated [21]. This paper addresses solely the software architecture, defined as the high-lev el desig n of the system arch itecture and the result of made design decisions [23,2 0,21] . The architecture is derived from the FRs and NFRs in an iterative design process t hat includes its design, evaluation and modification [23,24,20] . The design process is determin ed by reusability, e.g. architectural patterns, methods of overcoming the ga p between architecture and requirements, e. g., DDD , and intuition, e.g., architect exp erience [23,24] . High-level design dec isions relate to FRs and are used along the further development of th e architecture. Our d esign decisions are : • Usage of Retriev al-Augmen ted Generation (RAG) pattern and LLM dom ain adapters: RAG was chosen as it allo ws to add real-time and updated knowledg e to an LLM [2 5,26]. Adapter s allow a domain specific param etric efficien t fine -tunin g of an LLM [27,28]. RAG systems in combination with fine- tuning outperform RAG systems without any fine -tuning [29] . The combination of both approaches shall increase the domain spec ific information in the output of the LLM. • Usage of a co nversational agent: The textual and linguistic use of the co gnitive assistant ( CA ) should be enabled by a conversation al agent b ased on the conce pt presented in [ 30]. • Generalization of Security Checks: Security checks are requir ed to counteract potentially harmful content, such as inco rporated user feedb ack. This is d ue to the fact th at RAG systems are susceptible to malicious con tent in their database [31] . DDD ap proach es th e dev elopment o f c omplex software by focusing on one core domain and defining multip le bounded contexts with their own ub iquitous lan guages [32]. Bounded contexts describe the context within a model where boundaries are clearly d efined [33] . To derive boun ded contexts from FRs the English inf ormal strategy , firstly introduced by Abb ott [34] , was used. This strategy systematically analyzes textual description of r equirements by mappin g nouns to possible objects, verbs to p ossible procedures and adjectives to possible properties. The extracted word s are th en used to create a domain model. A domain model characterizes the problem space [35] . Despite this analysis, real- world kn owledge and domain expertise is still required to crea te such a mod el, as it i s not enough to solely rely on the an alysis [34] . Having a unified domain model fo r an entire sy stem is not suitab le . Thus, domain partitioning into m ultiple bound ed contexts , which should be un ified, is necessary [32]. On th at b asis, the following bounded contexts were der ived through the analy sis of the FRs co mbined with the h igh -level design d ecisions an d the resulting p arti ti on ing of the domain model : Chat, Retriever, LLM, Knowledge base, Feedback, Conversational A gent, User. To further design the software architectu re, architectural design options must be selected and proper ties and relationships of software componen ts mu st be defined [23] . The d iscovered bounded contexts wer e used as the foundation for the design of the software architecture . Based on the properties of bounded contexts, suggestions exist to implement each bo unded context as a microservice, creating a modular micro service architecture [36] . Micr oservices are cohesive, ind ependent processes that are lo osely coup led, independently d eployable, and communicate v ia messages [37 – 39] . A MAS decomp oses the system into m ultiple auton omous entities, such as ag ents or LLM -based agents, that work together to achieve the requirements of a system [40,21]. LLM- based agents enhan ce the capabilities of LLMs by integratin g interactive problem-solv ing functions, so called function calling. They have ex panded the range o f prob lems addressable by LLMs, including those that require external information through inter action with software tools [41,40 ]. The resu lting architecture design is a micr oservice-b ased MAS, where the architectur al components are microservices that simultaneously function as agen ts. Such architecture is flexible, scalable, and interchangeab le. 4 After hav ing develo ped a first version of the architecture, the next phase consist ed o f its design evaluation [23] by focus groups . Focus group is a qualitative research method that gathers a small, diverse g roup of in dividuals to discuss specific topics, aiming to explore their attitu des, p erceptions, beliefs, and opinions throug h g uided in teraction. [42] Sev en d omain experts participated in two o nline focus g roups, apart one month f rom each o ther, to assess aspects of the proposed architecture . Th e domain experts stemmed fro m acad emia and brought different levels of e xperien ce in research, software developm ent, QMS, and LLMs. In the first focus group, we presented the arch itecture to t he m wh o focused on id entifying strengths, limitations, and areas for impr ovement, ensuring alignment with user needs and industry stan dards. Based on their feedb ack, we impro ved the architecture, which was again presented to th e same s even domain experts in the second focus group. They an alysed and provided feedback to the second version, which was u pdated ba sed on the d iscussion. At th e end of the second focus group, a final version of the architecture was presented an d accordingly accepted by the participants. 4. Results The following chapter presents the resulting design of the architecture, sh own in Figure 1 . A brief overv iew of the architecture , the relationships between its components, and their description s is provided . Collaboration and comm unication are orchestrated by architectural compon ent ChatController . It also ser ves as th e interface to the user in terface ( UI ). Each core functionality is represented by a controller , which manages its interactio ns with the other controllers. The textual and voca l usage is orchestrated b y Conversationa lAgent Controller through th e interaction between the architecture components ConversationalAg ent , LLMAgen t and RAGRetrieval . To incorporate user feedback into the system, FeedbackEvaluation Figure 1: Architec ture of the LLM-Co gnitive Assistant 5 Controller manages the interaction between FeedbackEvalua tion and RAGRetrieval . The controller for ConversationHistory is solely r esponsible for providing this co mponent. User s may be assigned to seve ral user g roups, with d ifferent authorizations, which are managed through an Access Control List (ACL). The component ConversationalAg ent enables verbal and textual in teraction with th e cog nitive as sistant, ensuring a multimodal system with at least two modalities. It is based o n the conv ersational agent architecture proposed by [ 30] . Componen t I nput Handling converts supported input formats into a textual representation, whereas O utput Handling converts the resp onse of the system into the needed output format, i.e., text or v oice. The co mponents Dialog Ma nager, Natural Language Un derstandin g (NLU) and Natural Languag e Generation (NLG ) are realized by the LLMAgen t , as indicated by the grey background of the componen ts in the architecture d iagram . Architectural component LL MAgent is an LLM -based agent, which co mprises two components: Guardrailling and LLM . Guardrailling is a safety measurement to regulate user interactio ns with LL Ms with the purpose of ensuring that the LLM adheres to ethical and organization principles, e.g., complian ce guidelines based on government regulations and employee equality policies. It is r ealized b y identifying harmful conten t in user pr ompts and m odel responses [43,44] . The second component contains an LLM with mu ltiple do main adapters. Adapters allow to fine tu ne a general pu rpose LLM for a d omain specific task, while still using the capab ilities of an LLM without requiring extensive computation al reso urces [27,45] . An LLM chosen for this component needs to provide the functionality o f function calling to enable th e workflow of an LLM - based agent . Function ca lling further enables users to use m athematical functions needed for p rocess-specific ca lculations an d process data analysis that wou ld serv e as a first step to empower users to perform continuous improvem ent , e.g., optimization of product setup tim e. Architectural co mponent RAGRetrieval m anages and retrieves documents nec essary fo r injecting an LLM with up - to -date kno wledge. Do cuments can be loaded from a defined source, e.g., file system, as well as directly updated, e.g., by a feedback tick et . Do cuments are first analy sed, considering aspects such as layou t and table structure, and subsequently unified into a comm on format based on the concept presented in [4 6]. Afterwards, they are chunked, em bedded and stored in a vecto r store. A second compon ent utilizes the r etrieval of document chunk s for a given query. Architecture compon ent F eedbackEvaluation enables the functionality to continuously impro ve th e systems knowledge through users . User s can either flag an answer as “insu fficient” for an incorrect answer, or as “exten d” for a par tially complete answer, cr eating a feedback ticket. Specific user gro ups are allowed to r ewrite such a nswers, e. g. , user g roup k -1 (supervisors) , and extend a feedb ack ticket with additional documents or update ex isting d ocuments. Feedbac k tickets undergo two security check s before b eing integrated into the vector store . T hey consist of a jailbr eak check and a f act check. The jailbreak checker verifies whether the feedback ticket does not co ntain adversarial input th at cou ld ch ange the behavior of the LLM, e.g. jamming attac ks [31]. The fact check er ensures that injected knowledge is within the scope of co ntext provided by the feedback ticket. It is possible for both c h ecks to b e performed by hum ans, by an LLM , or by th eir combination . To enable continu ous learning, feedback tickets are stored and ca n be retrieved by specific u ser gro ups, e.g. u ser group k at the managerial level. These tickets can provide anonymous insights in to p otential u ser knowled ge gaps, e.g., lack of data analytics skills, and system response quality, e.g., rate of incomplete answers . Ar chitectural component ConversationHistory stores and pr ovides all user’s conversation s for resumption . 5. Conclusions The f indings of this study highlight the potential of LLM - CAs to enhance operational p erformance in manufacturing within QMS. By integr ating artificial intelligence -driven capabilities, LLM-CAs can facilitate decision -making, continuous impr ovement, and kno wledge management while aligning with QMS requirements. Existing LLM - CA s research does not focus on QMS, wh ich can be fundamental for the integration of such a technology into manufacturing [1]. Th is research addresses th is gap by proposing a h olistic software architectur e for LLM-CAs within QMS. The pro posed solution integrates fine -tuning , RAG, hierarchical k nowledge updates, and co mpliance mechanisms, ensuring ef ficient, auditable, and human -centred knowledge managemen t and continuous improvement . On the o ne hand, the architecture enables emp loyees to be key player s within work activities, e.g. , performing process impr ovements . On the other hand, it supports them b y means of technological function alities, e.g., data analysis, versio n control of documents, and compliance checks. The limitation of this study lies on its theoretical developm ent, assuming therefore its full fu nctionality without previous testing. On that basis, our f uture research will operationalize an d evaluate the proposed architectur e based on real m anufacturing use cases an d u ser evaluation . Mor eover, different available LLMs will be tested for performance. Therefore, practical r esults can lead to an update of the current proposed architectur e that would refine it and enhance its applicability in manufacturing. Acknowledgem ents This research was developed in the context of the project ZUKIPRO [ZUK- 1-000 5], which is funded by the German Federal Ministry of Labour and Social Affairs (BMAS) and the European Union via th e European So cial Fund Plus (ESF Plus) as part of th e ‘Future Centres’ prog ramme. 6 References [1] Galdino, M., Hamann, T., Ab delrazeq, A., I senhardt, I., 2025. Large Language Mo del ‐ Based Cognitive A ssistants for Qualit y Management Systems in Manufa cturing: A Re quirement Analys is. Engineering Reports 7 (10). [2] Kernan Fre ire, S., 2025. LLM-Pow ered Cognitive Ass istants for Knowledge S haring among Fa ctory Operators, 179 pp. [3] Kernan Fre ire, S., Foosheri an, M., Wang, C. , Niforatos, E., 2023. Harnessing Lar ge Language Mo dels for Cognitiv e Assistants in Factories, in: Proce edings of the 5 th International Co nference on Conversational U ser Interfaces. CUI '23: ACM conference on Conversational User I nt erfaces, E indhoven Nether lands. 19 07 20 23 21 07 2023. ACM, New York, NY, USA , pp. 1 – 6. [4] Ding, Q., Di ng, D., Wang, Y., Guan, C., Di ng, B., 2024. Unra veling the landscape of lar g e language m odels: a system atic review and f uture perspectives. JE BDE 3 (1), 3 – 19. [5] Ainapure, A ., Dhamane, S., Dhage, S. Embodied Epistemology: A Meta-Cognitive Ex ploration of C hatbot-Enabled Docume nt Analysis. International C onference on Ev olutionary 202 3, 1 – 6. [6] Minglana, J ., Tobias, R.R., R oxas, R.E. Art ificial Intelligence Applications i n Quality Manageme nt System: A Bibliometric Stu dy, 947 – 952. [7] International Organization for Standardizat ion. ISO 9000: 2015 (E): Quality mana gement systems - Requir ements Geneva, 2015. [8] Zhao, W. X . et al., 2023. A Sur vey of Large Lan guage Models, 140 pp. http://arxiv.org /pdf/2303.1822 3v15. [9] Vaswani, A ., Shazeer, N., Parmar , N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L. , Polosukhin, I . , 2017. Attention I s All You Need, 1 5 pp. http://arxiv.org /pdf/1706.0376 2v7. [10] Fan, H., Liu, X., Fuh, J.Y.H., Lu, W.F., L i, B., 2025. Embo died intelligence in ma nufacturing: lev eraging large language models f or autonomous ind ustrial robotics. J I ntell Manuf 36 (2), 1141 – 1157. [11] Kernan Fre ire, S., Wang, C., Foosherian, M., W ellsandt, S., Rui z - Arenas, S., Nif oratos, E., 202 4. Knowledge shari ng in manufacturi ng using LLM-pow ered tools: user stu dy and model benchmarking. Frontiers in ar tificial intellige nce 7, 1293084. [12] Bendoly, E. , Chandrasekara n, A., Lima, M.d. R.F., Handfield, R., Khajavi, S.H., R oscoe, S., 202 4. The role of gen erative design a nd additive manuf acturing capabi lities in developin g human – AI symbiosis: Evidence from m ultiple case stu dies. Decision Sci ences 55 (4), 325 – 345. [13] Hoyle, D., 2 018. ISO 9000 qu ality systems ha ndbook: Increasi ng the quality of an org anization's out puts, Seventh e dition ed. Routled ge Taylor & Franc is Group, London, N ew York, 1 on line resource. [14] Fonseca, L., D o mingues, J.P. , 2017. ISO 900 1:2015 EDITI ON- Management, q uality, and value. I nternational Journal for Qua lity Research; Vol 1 1, No 1, 2017. I SSN 1800-6450. [15] Solomon, N.P. , Bester, A., Moll, C.M., 201 7. Diffusion of a q uality management syste m: a case stu dy. SAJI E 28 (2). [16] Mustroph, H ., Rinderle-Ma, S. Des ign of a Qualit y Management Syst em Based on the EU AI Act. Legal Knowledge and I nformation Syste ms 2024. [17] Fadilasari, D .P., Roy Ghata k, R., Garza-Reyes, J.A., Joshi, R., Kandasamy, J., 2024. Adopting qua lity manageme nt practices in t he industry 4.0 era: an investigation into the challenges. T otal Quality Management & B usiness Excelle nce 35 (9-10), 1098 – 1123. [18] Murmura, F., Musso, F., Bra vi, L., Pierli, G., 2024. The ro le of quality management syste ms in fostering the internati onal competitiveness of companies. I JQRM 41 (7), 1979 – 199 9. [19] Bucaioni, A., Weyssow, M., He , J., Lyu, Y., L o, D., 2025. A Fu nctional Software Refere nce Architectur e for LLM-Integrated S ystems, 5 pp. http://arxiv.org /pdf/2501.1290 4v1. [20] van der V en, J.S., Jansen, A.G.J ., Nijhuis, J.A.G. , Bosch, J. Desig n Decisions: The Bridge between Rationale and Arc hitecture, in: , pp. 329 – 348. [21] Weyns, D., H elleboogh, A., St eegmans, E., W olf, T. de, Merte ns, K., Boucké, N., Hol voet, T., 2004. A gents are not part of the problem, agents can sol ve the problem, in: Pro ceedings of t he OOPSLA Workshop on A gent-Oriented Metho dologies, 2004, pp. 101 – 112. [22] Lapouchnian, A., 2005. Goal-orient ed requireme nts engineering: An overview of the c urrent research . University of T oronto 32. [23] Falessi, D., Ca ntone, G., Kazm an, R., Kruchte n, P., 2011. De cision - making techniq ues for softwar e architecture desi gn: A comparati ve survey. ACM C omputing Surve ys (CSUR) 43 (4), 1 – 28. [24] Kruchten, P., 1995. Mommy, wh ere do software architectures c ome from, in: Proceed ings of the 1st I n tl. Workshop o n Architectures f or Software Systems , pp. 198 – 205. [25] Arslan, M., Gh anem, H., Munawar , S., Cruz, C ., 2024. A Sur vey on RAG with LL Ms. Procedia Compu ter Science 2 46, 3781 – 3790. [26] Ovadia, O., Brief, M., Mishae li, M., Elisha, O., 2023. Fine-tuning or retrieval? com paring knowledge injection in llms . arXiv preprint arXiv:2312.0593 4. [27] Hu, Z., et a l., 2023. Llm-adapters: A n adapter family for parame ter- efficient fine-t uning of large langua ge models. arX iv preprint arXiv:2304.0193 3. [28] Setty, S., Tha kkar, H., Lee, A., Chung, E., Vi dra, N., 2024. I mproving retrieval for ra g based questio n answering models on financial documents. arXi v preprint arXi v:2404.07221. [29] Nguyen, Z., e t al., 2024. Enha ncing Q&A with D omain -Specifi c Fine- Tuning and I terative Reasonin g: A Comparative Study. arXiv pre print arXiv:2404.1179 2. [30] Colabianchi, S., Tedeschi, A., Costantino, F., 2023. Human-technology integration with industrial conv ersational agents: A conceptual architecture and a taxonomy f or manufacturi ng. Journal of I ndustrial Information I ntegration 35, 10 0510. [31] Shafran, A., Sc huster, R., Shm atikov, V., 2024. Machine against t he rag: Jamming retrieva l-augmented gen eration with bloc ker documents. arXiv preprint arXiv: 2406.05870. [32] Evans, E., Des ign, D.-D., 2003. T ackling comple xity in the heart of software. Eric Ev ans ‐ 1st Edition-Addis on-Wesley Professi onal. [33] Evans, E., 20 14. Domain-driven des ign referen ce: Definitions an d pattern summ aries. Dog Ear P ublishing. [34] Abbott, R.J., 1983. Program de sign by informal E nglish descript ions. Communications of the AC M 26 (11), 882 – 894. [35] Hasselbring, W ., 2018. Softw are architecture: P ast, present, f uture. The essence of sof tware engineerin g, 169 – 184. [36] Kapferer, S., Zimmermann, O ., 2020. Doma in -specific L anguage and Tools for Strat egic Domain-driven D esign, Context Mapping and Bounded Context Modeling, in: MOD ELSWARD, pp. 299 – 306. [37] Dragoni, N., Gi allorenzo, S., L afuente, A.L., Mazzara, M., Mon tesi, F., Mustafin, R., S afina, L., 2017. M icroservices: yes terday, today, an d tomorrow. Prese nt and ulterior s oftware engineer ing, 195 – 216. [38] Lewis, J., Fow ler, M., 2014. M icroservices: a d efinition of this n ew architectural term. MartinFowl er. com 25 (14-26), 12. [39] Richards, M., 2015. Software architecture patter ns. O’Reilly Med ia, Incorporated 10 05 Gravenstei n Highway Nort h, Sebastopol, CA … [40] Jin, H., Huang, L., Cai, H., Ya n, J., Li, B., Chen, H. , 2024. From llms to llm -based agents for software enginee ring: A survey of current, challenges and fu ture. arXiv pre print arXiv:2408.02 479. [41] Abdelaziz, I ., et al., 2024. Gr anite-function callin g model: Introduc ing function callin g abilities via mul ti -task lear n ing of gran ular tasks. arXiv preprint arXiv: 2407.00121. [42] Then K. L., Rankin J. A., Ali E., 2014. Focus gr oup research: w hat is it and how can it be used? Can J Car diovasc Nurs. (Wi nter;24(1)), 16 – 22. [43] Dong, Y., et al., 2024. Safeg uarding large lan guage models: A sur vey. arXiv preprint arX iv:2406.0262 2. [44] Padhi, I., et a l., 2024. Granite G uardian. arX iv preprint arXiv:2412.0772 4. [45] Mathav, R.J., VM , K., Warrier, H . , Gupta, Y., 2 024. Fine tuning l lm for enterprise: Practi cal guidelines and recommen dations. arXiv preprint arXiv:2404.1077 9. [46] Livathinos, N ., et al., 2025. Do cling: An Efficie nt Open-Source Toolkit for AI-driven D ocument Conversi o n. arXiv prepri nt arXiv:2501.1 7887.