CREATING A VIRTUAL ASSISTANT FOR THE CASE OF APICULTURE

CREATING A VIRTUAL ASSISTANT FOR THE CASE OF APICULTURE S. Karetsos, I.T. Karageorgi, K. Demestichas, C. Costopoulou Agricultural University of Athens (GREECE) Abstract The digital transformation of agriculture increasingly relies on artificial intelligence (AI) to deliver personalized, efficient, and context-specific support for practitioners. This study presents the design, fine-tuning, and empirical evaluation of a domain-specific AI chatbot developed to enhance both educational and professional activities in Greek apiculture. Drawing on a certified knowledge base constructed from authoritative local and international beekeeping resources, the chatbot was locally deployed using an open-source platform and refined through iterative enrichment of domain-specific content. Performance assessment involved structured testing with both experienced beekeepers and agricultural sciences students, using a diverse question set to evaluate accuracy, clarity, and pedagogical value. Results indicated marked improvement in fact-based and routine management queries after corpus enrichment, particularly in multiple-choice and true-false question types, while short-answer performance remained constrained by limitations in language modeling and data structure. User feedback highlighted the tool’s value in clarifying key concepts, supporting independent learning, and facilitating rapid access to certified information, with the greatest impact seen among novices and in routine management scenarios. Challenges related to translation, corpus structuring, and the absence of native Greek language support were identified, alongside the need for further technical optimization and broader integration into educational curricula. This work demonstrates the practical and educational benefits of targeted chatbot training in apiculture and underscores the potential of AI-driven virtual assistants to promote sustainable, knowledge-based development in rural agricultural sectors. Keywords: Chatbots, Virtual assistants, Farmer training, Education technology, Beekeeping, Agricultural technology, Natural Language Processing, Decision Support Systems. 1 INTRODUCTION The ongoing digital transformation in agriculture is reshaping how traditional practices adapt to technological innovations, particularly in knowledge-intensive sectors such as apiculture. Among these innovations, virtual assistants, including chatbots powered by artificial intelligence (AI), have emerged as versatile tools that automate communication, streamline decision-making, and offer personalized, timely support to practitioners in various agricultural domains [1]-[3]. The integration of such technology is especially pertinent to beekeeping, a practice that is deeply embedded in both the economy and culture of Greece. With approximately 2.1 million hives and an annual honey production fluctuating between 15,000 and 20,000 tons, Greek apiculture represents one of the most vital agricultural activities, not only contributing substantially to local and export markets but also supporting biodiversity through pollination [4]. Despite the sector’s significance, Greek beekeepers frequently encounter critical barriers, including limited educational and training resources on hive management, disease diagnosis, and reliable information access. The rapid evolution in beekeeping methods, sparked by environmental challenges, regulatory changes, and market demands, underscores the need for continuous, accessible training mechanisms for both novice and experienced apiculturists. In this context, AI-powered chatbots offer a promising solution, serving as interactive digital assistants capable of delivering certified knowledge, practical guidance, and reinforcement for self-directed learning irrespective of time or location [5]. Current research in agricultural informatics has extensively examined the application of chatbots for several topics in agriculture, yet their potential in supporting beekeepers with domain-specific queries remains underexplored [1]-[3]. More specifically, research shows that chatbots, powered by artificial intelligence and large language models, are increasingly deployed in agriculture for tasks such as crop monitoring, disease diagnosis, resource management, and providing technical support. Most chatbot applications have focused on crop selection, pest management, irrigation, and advisory services for farmers, using natural language processing and deep learning to deliver personalized, real-time responses and decision support. Chatbots have demonstrated improvements in efficiency, sustainability, and extension outreach, especially for farmers in remote or underserved regions [5]. Literature reviews reveal numerous studies on chatbots for agricultural crops, poultry farming, and plant disease detection, but limited results exist regarding their application in beekeeping, highlighting a significant research gap [6]-[7]. In other words, the apiculture sector remains underexplored, both internationally and within the EU context. Studies affirm that success in agricultural chatbots depends on robust domain-specific training, multilingual functionality, and continuous validation with users and experts to ensure reliable and context-aware responses [8]-[9]. Deep learning and natural language processing are emphasized as critical technologies that transform chatbots from simple automation tools to sophisticated systems capable of interpreting complex queries and supporting learning, emotional engagement, and adaptation to individual needs [10]-[11]. Research stresses the importance of combining open-source models and rich, certified datasets to address specialized sector requirements and language barriers, especially in local languages like Greek [12]- [13]. It also underscores the growing role of AI chatbots in agriculture, the need for targeted development for apiculture, and the potential for tailored, knowledge-driven digital assistants to transform training and support in beekeeping. Thus, in a nutshell, most existing studies are limited by language barriers, insufficient domain-specific training data, or lack of evaluation in real-world contexts. The present research positions itself within this knowledge gap by investigating the development, deployment, and assessment of a specialized chatbot tailored to the educational and decision-support needs of Greek beekeepers. The objectives of this study are: • To construct a robust, certified knowledge base for apiculture from authoritative sources and national guidelines; • To fine-tune a large language model for precise, context-aware responses to beekeeper queries; and • To evaluate the chatbot across different user groups—experienced practitioners and agricultural students, by measuring accuracy, user satisfaction, and adaptability to varying expertise levels. Contributing to methods for AI-supported training and information access in apiculture, this work seeks to bridge gaps in rural education infrastructures, promote sustainable beekeeping practices, and contribute to a more resilient, knowledge-driven agricultural sector in Greece and beyond. 2 METHODOLOGY The employed research methodology consisted of the following phases: 1. Knowledge Base Construction The initial phase centered on assembling a robust, certified knowledge base specifically for apiculture. Content was curated from authoritative sources, including: • Textbooks and guides provided by the Beekeeping Laboratory of the Agricultural University of Athens. Specifically, the corpus consisted of 17 documents in PDF format. • Internationally recognized manuals such as Advanced Beekeeping (Purdue University) [13], Beekeeping Basics (Penn State College) [14], First Lessons in Beekeeping (Dadant) [15], and The Beekeepers Handbook (by Sharma et al.) [16]. • These documents encompassed key topics: hive management, disease prevention/treatment, equipment, nutrition, seasonal interventions, and honey production. All documents – originally in Greek – were translated into English, primarily using the DeepL engine, to ensure language compatibility with the selected language model. The knowledge base was designed for exhaustive coverage and targeted granularity, balancing technical depth with accessibility for novice users. 2. Model Selection and Technical Implementation Based on a comparative review of available open-source models, the GPT4ALL platform [17] was chosen for its local execution capability, privacy assurances, and user-friendly interface. Within GPT4ALL, the Llama 3.2 3B Instruct model (Meta) was selected for the following reasons: • Multilingual support and strong context retention. • Efficient instruction-following and summarization tasks. • Technical compatibility with Windows, MacOS, and Linux systems. • Alternative models such as Phi-Mini Instruct and Qwen 2.5 Instruct (Microsoft, Alibaba) were also evaluated for comparative purposes (see Table "Available Models" in thesis), but Llama 3.2 was prioritized for its performance and resource efficiency under local deployment. 3. Corpus Integration and Model Fine-Tuning The curated corpus was manually uploaded to the GPT4ALL environment as text files, facilitating direct ingestion and model fine-tuning at the user level. The translation process and corpus structuring were essential given the absence of direct Greek language support, ensuring semantic fidelity and organization. Supplementary material was incrementally added through a second phase, allowing sequential improvement in domain knowledge and model adaptability. The model was trained to respond to targeted queries across the knowledge base, simulating real-world dialogue scenarios encountered by Greek beekeepers. 4. Evaluation Approach The chatbot was evaluated through an empirical protocol using two discrete user groups: • Experienced Beekeepers: Assessed accuracy, reliability, and technical soundness of chatbot responses, with questions partitioned into multiple-choice, true-false, and open short-answer formats. • Agricultural Sciences Students: Evaluated accessibility, clarity, pedagogic support, and selflearning utility. • A set of 20 control questions (8 multiple-choice, 9 true-false, 3 open-ended) was first posed in Greek and then translated into English for chatbot input. Responses were translated back to Greek for validation by domain experts and novices. Success rates were calculated as the proportion of correct answers per question type, benchmarked against results for ChatGPT, Gemini, and Copilot AI tools. 5. Supplementary Analysis and Comparative Testing During the development cycle, four additional comprehensive apiculture guides were uploaded to the model for the second round of testing. This allowed assessment of incremental improvements in chatbot accuracy, particularly in question categories related to advanced disease identification and hive management. 6. Data Security and Local Deployment All training and evaluation were conducted through local execution to safeguard data privacy and support modeling in environments with restricted or unreliable internet access. GPT4ALL’s native capacity for offline operation and local data storage reinforced participant anonymity and compliance with ethical research practices. Notably, the methodology was guided by sectoral best practices in agricultural informatics and AI for education, with continuous reference to technical and pedagogical literature on chatbot fine-tuning and evaluation. Translation was both a technical and conceptual constraint, requiring human oversight and semantic review to avoid information loss. 3 RESULTS The following paragraphs present the results of the performed research activities. 1. Performance of the Apiculture Chatbot Evaluation of the specialized chatbot tailored for beekeeping was conducted in two phases – initially with the basis corpus (GPT4ALL 1), and subsequently with the enriched corpus (GPT4ALL 2). The chatbot was tested on a standardized set of 20 questions (8 multiple-choice, 9 true-false, 3 development/short-answer), administered to two representative user groups: experienced beekeepers and undergraduate agricultural students. First Evaluation Phase (Core Knowledge Base): • Multiple-Choice Questions: 3 out of 8 answered correctly (38% success rate) • True-False Questions: 4 out of 9 answered correctly (44% success rate) • Short-Answer Questions: 1 out of 3 answered correctly (33% success rate) Compared to cloud-based systems (ChatGPT 3.5, Gemini, Copilot), the initial chatbot lagged in accuracy—commercial models averaged 75–87.5% in multiple-choice and 55–66.7% in truefalse categories. Notably, performance for short-answer questions was similarly low across all systems (33%), indicating consistent difficulty in high-context queries. Second Evaluation Phase (Expanded Training Corpus): • Four additional authoritative apiculture guides were integrated. • Multiple-Choice Questions: Improved to 5/8 correct (62.5%) • True-False Questions: 5/9 correct (55.5%) • Short-Answer Questions: 0/3 correct (0%) Thus, targeted corpus enrichment led to marked gains on structured, factual queries (multiple choice, true-false), but the tool’s ability with open-ended responses remained limited. 2. User Group Feedback • Experienced Beekeepers: Reported strong technical accuracy and reliability for mainstream management topics (e.g., hive inspections, seasonal feeding, honey harvesting, terminology clarification). Recognized value in rapid information access, but underlined the necessity for human expert input in advanced disease diagnosis and nuanced hive interventions. • Agricultural Sciences Students: Valued the chatbot for clarity, learning support, and interactive practice. It enhanced classroom education by enabling self-testing, immediate feedback, and exploration of simulated scenarios. 3. Limitations & Technical Challenges • Language: No native Greek language support; all content required translation, causing occasional semantic drift. • Corpus Structure: Chatbot struggled with large, unstructured information, performing best with concise, thematic datasets. • Short-Answer Complexity: Difficulty with interpretive, synthetic queries and with offering creative solutions to open-ended problems. • Offline Operation: Local deployment ideal for privacy/field settings, but lacking continuous model improvements and cloud-based refinements. 4. Summary Comparative Analysis Enriching the chatbot’s dataset with targeted apiculture references distinctly improved accuracy for factual and objective queries, demonstrating the critical impact of domain-specific data. Nonetheless, as depicted in Table 1, performance remained below commercial alternatives for complex question types, confirming that ongoing domain adaptation and technical optimization are needed for broader application in beekeeping education and practice. These results highlight the real-world progress and persistent challenges of deploying AI chatbots in apiculture. Fine-tuned, domain-rich datasets yield strong advances, but language, data structure, and model limitations continue to constrain the scope of effective, reliable support—reinforcing the need for ongoing research and refinement. Table 1. Comparative performance of selected AI-powered chatbots in apiculture education. Success Rate (as percentage) Artificial Intelligence Applications (Chatbots) True/False Questions Multiple Choice Questions Open-ended (short answer) Questions Cloud-based chatbots Copilot 55.60 87.50 66.66 Gemini (Bard) 66.70 75.00 66.66 ChatGPT 3.5 66.70 87.5 33.33 Local specialized chatbots GPT4ALL (Llama 3.2 1B Instruct) (1) 44.00 38.00 33.00 GPT4ALL (Llama 3.2 1B Instruct) (2) 55.50 62.50 0 4 CONCLUSIONS This study demonstrates that domain-specific, AI-powered chatbots have substantial potential to enhance both professional practice and education in apiculture, especially when developed and finetuned using certified, highly targeted beekeeping knowledge. Through the design and two-phase empirical evaluation of a locally deployed, open-source chatbot tailored to Greek beekeepers’ needs, the research identified both notable strengths and persistent challenges for real-world deployment. The chatbot’s utility was most evident in its ability to address fundamental aspects of hive management, such as routine inspections, feeding, and honey harvesting. It also proved particularly effective in clarifying terminology, supporting step-by-step procedures, and reinforcing theoretical understanding. Novice users acknowledged the tool’s value as a digital learning companion, offering immediate feedback and an accessible means to supplement formal coursework through problem-solving in simulated scenarios. Professional beekeepers found the chatbot to be a convenient reference for certified information in daily practice, though they continued to rely on human expertise for more complex decisions, such as disease diagnosis and intervention planning. The study specifically highlights the impact of targeted corpus enrichment, showing that augmenting the chatbot’s training set with additional authoritative texts measurably improved performance in structured query types, including multiple-choice and true-false questions. These findings underscore the critical importance of focused, high-quality data when fine-tuning AI applications for agriculture. However, challenges remain with respect to language support and technical adaptability. The absence of native Greek functionality necessitated repeated translation between Greek and English, which at times led to semantic inaccuracies and diminished response fidelity. Furthermore, while the chatbot performed well with concise, thematically organized data, it struggled with unstructured, open-ended, or highly complex queries (an area where leading commercial AI models retain a clear advantage). Despite these limitations, the research affirms the role that AI chatbots can play in bridging educational and informational gaps, particularly in rural or resource-constrained contexts where traditional training and extension services may be lacking. Such digital tools have the capacity to foster digital literacy, promote lifelong learning, and catalyze the modernization of knowledge exchange in beekeeping. The offline, privacy-preserving architecture of the tested system further enhances its relevance and applicability in the field. Looking forward, several avenues for improvement are evident. Expansion of the training corpus, integration of native Greek language processing, ongoing technical optimization, and more seamless incorporation of expert knowledge will be essential to maximize both accuracy and user engagement. Moreover, broader integration within agricultural curricula and vocational programs could unlock even greater educational value and sectoral impact. In summary, the present work establishes a strong foundation for the use of chatbots in apiculture, while at the same time highlighting the ongoing need for advanced linguistic support and participatory, domain-driven development. With continued research and iterative refinement, AI-powered virtual assistants have the potential to become indispensable resources for sustainable and knowledge-driven beekeeping in Greece and beyond. ACKNOWLEDGEMENTS This publication is part of the TALLHEDA project that has received funding from the European Union's Horizon Europe research and innovation programme under grant agreement No. 101136578. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Executive Agency (REA). Neither the European Union nor the granting authority can be held responsible for them. REFERENCES [1] Chatrabhuj, K. Meshram, U. Mishra, and U. Rathnayake, “Application of Artificial Intelligence in Agri-Tech, Environmental and Biodiversity Conservation,” Array, pp. 100412–100412, May 2025, doi: https://doi.org/10.1016/j.array.2025.100412. [2] M. T. Kuska, M. Wahabzada, and S. Paulus, “Ai-Chatbots for Agriculture - Where Can Large Language Models Provide Substantial Value?,” SSRN, Jan. 2024, doi: https://doi.org/10.2139/ssrn.4685971. [3] P. D. Reddy, S. Sampath, P Jayanth, Bhanu Prakash Kakarla, and R. M. Balakrishnan, “Agri Assist: An AI Integrated Farmer Assistant,” Procedia Computer Science, vol. 258, pp. 3510– 3522, Jan. 2025, doi: https://doi.org/10.1016/j.procs.2025.04.607. [4] Hellenic Statistics Authority (ELSTAT), “ANNUAL AGRICULTURAL STATISTICAL SURVEY: Year 2023”, Statistics.gr, June 2025, Retrieved from https://www.statistics.gr/documents/20181/bbb985c3-50e5-1829-7aec-71acb19c175b [5] A. Ahmad et al., “AI can empower agriculture for global food security: challenges and prospects in developing nations,” Frontiers in Artificial Intelligence, vol. 7, Apr. 2024, doi: https://doi.org/10.3389/frai.2024.1328530. [6] P. K. Astuti, B. Hegedűs, A. Oleksa, Z. Bagi, and Szilvia Kusza, “Buzzing with Intelligence: Current Issues in Apiculture and the Role of Artificial Intelligence (AI) to Tackle It,” Insects, vol. 15, no. 6, pp. 418–418, Jun. 2024, doi: https://doi.org/10.3390/insects15060418. [7] M. Alleri et al., “Recent developments on precision beekeeping: A systematic literature review,” Journal of Agriculture and Food Research, vol. 14, p. 100726, Dec. 2023, doi: https://doi.org/10.1016/j.jafr.2023.100726. [8] Ξ. Kpodo and A. Pouyan Nejadhashemi, “Navigating challenges/opportunities in developing smart agricultural extension platforms: Multi-media data mining techniques,” Artificial Intelligence in Agriculture, Apr. 2025, doi: https://doi.org/10.1016/j.aiia.2025.04.001. [9] J. Khanifar, “Evaluating AI-generated responses from different chatbots to soil science-related questions,” Soil Advances, vol. 3, p. 100034, Jun. 2025, doi: https://doi.org/10.1016/j.soilad.2025.100034. [10] M. Nautiyal et al., “Revolutionizing agriculture: A comprehensive review on artificial intelligence applications in enhancing properties of agricultural produce,” Food Chemistry: X, vol. 29, p. 102748, Jul. 2025, doi: https://doi.org/10.1016/j.fochx.2025.102748. [11] T. H. Sambo, “NLP-Based Multilingual Chatbots for Farmer Advisory Systems,” Zenodo, Sep. 2025, doi: https://doi.org/10.5281/zenodo.17136502. [12] M. S. Allahyari, Sinisa Berjan, Hamid El Bilali, T. B. Hassen, and Soroush Marzban, “Assessing the Use of ChatGPT among Agri-Food Researchers: A Global Perspective,” Journal of Agriculture and Food Research, pp. 101616–101616, Dec. 2024, doi: https://doi.org/10.1016/j.jafr.2024.101616. [13] N. Carroll, et al. Advanced Beekeeping. Purdue Extension Publication 4-H-1059-W, Purdue University, 2022. [14] M. Frazier and E. Tomaszewski. Beekeeping Basics. Penn State Extension, College of Agricultural Sciences, Penn State University, 2024. [15] C. P. Dadant. First Lessons in Beekeeping. American Bee Journal, Hamilton, Illinois, 1918. [16] R. Sharma, R. Kiran Kumar, and P Vaijinath. The Beekeeper's Handbook: A Comprehensive Guide to Apiculture. Shineeks Deeper Knowledge Publishers. ISBN: 979-8-88940-982-3, 2023. [17] GPT4All – The Leading Private AI Chatbot for Local Language Models. Retrieved from https://www.nomic.ai/gpt4all