Artificial intelligence is redefining digital health by enabling applications to interpret natural language, process clinical context, and generate user-tailored guidance. ¹ Several studies have combined large language models (LLMs) with medical ontologies to support the identification of pathologies from clinical reports, capturing attributes such as type, severity, and location. ² Moreover, recent evidence suggests that prompting strategies can improve LLM performance in clinical reasoning tasks, ³ reinforcing their ability to interpret symptom descriptions and produce contextual responses that may support automated medical guidance tools. ^{1, 4} These advances are particularly valuable for making pharmaceutical knowledge more accessible to the general population in a responsible manner.

Common illnesses—such as influenza-like syndromes, cough, fever, and mild pain—represent a frequent reason for primary care consultation. For example, acute respiratory symptoms (e.g., cough and dyspnea) and upper respiratory infections were reported among the most frequent causes of consultation in a survey of primary care professionals (34.8% of responses). ⁵ At the same time, self-medication without professional supervision remains widespread and is associated with risks such as incorrect self-diagnosis, inappropriate treatment selection, incorrect dosage, harmful drug interactions, and masking of more serious conditions. ^{6, 7, 8} In this context, intelligent medication guidance systems may serve as informational tools to help users make safer decisions, particularly when dealing with over-the-counter (OTC) medicines.

Internationally, systems with similar objectives have been explored. For example, ⁹ developed a pharmaceutical conversational agent integrated with Micromedex to answer medication-related queries using natural language processing. In Europe, AI-supported clinical reasoning approaches have been studied to analyze pharmacological emergency data and support treatment decision-making. ¹⁰ In the Netherlands, intelligent systems have been shown to optimize medication alerts and recognize drug interactions using AI. ¹¹ Similarly, research in Egypt has reported advances in predicting drug interactions using deep learning models. ¹² However, many of these systems rely on complex infrastructures, closed models, or limited customization based on user profiles.

In Latin America, the development of AI-assisted pharmacological tools remains at an early stage. Regional analyses indicate that AI health projects are fragmented, experimental, and often limited to isolated applications rather than integrated solutions. ¹³ In Colombia, for example, ¹⁴ conducted a bibliometric review showing that health chatbots are expanding—especially in mental health and medical education—while highlighting the need for stronger scientific justification prior to large-scale deployment. Overall, few systems in the region integrate open and modular architectures that enable personalization based on user context, which can limit scalability in academic and low-resource settings.

In Peru, digital health has progressed in areas such as telemedicine and the adoption of information technologies for health services. Teleconsultation services expanded during the COVID-19 pandemic to improve access in areas with limited in-person coverage. ¹⁵ Nevertheless, institutional, regulatory, and structural challenges continue to hinder the full adoption of digital health solutions. ^{16, 17} Despite this progress, initiatives specifically focused on AI-assisted OTC medication guidance appear to remain limited. Although the present development was implemented in Peru, its modular web-based design makes it potentially applicable in other contexts with similar needs.

To address this gap, we propose an Intelligent System for the Recommendation of Medicines for Common Diseases using the Gemini API. This work focuses on the design and architecture of a web-based system rather than a clinical pilot implementation. The system processes user-reported symptoms and contextual information (e.g., age, allergies, and pre-existing conditions) and generates informational OTC medication recommendations and safety warnings using Google’s Gemini API. The system is intended for informational and educational purposes only and does not provide medical diagnosis or replace professional medical advice. Users are encouraged to consult a qualified healthcare professional, particularly when symptoms are severe, persistent, or unclear.

The proposed system emphasizes transparency, adaptability, and ethical design by incorporating strict structured prompting, server-side JSON/schema validation, and safety rules that suppress recommendations in high-risk scenarios. The main contributions of this work are: (i) a modular web-based architecture for symptom-driven OTC medication guidance, (ii) structured prompt engineering with schema-constrained JSON outputs for reproducibility, and (iii) safety-oriented warning generation and risk-based recommendation suppression.

In this section, we describe the methods used for the design, implementation, and operation of the Intelligent System for the Recommendation of Medicines for Common Diseases. The system is a web-based application that collects patient-reported symptoms and basic clinical context (e.g., age, sex, allergies, and pre-existing conditions), structures this information into a standardized JSON prompt, and uses the Gemini API to generate informational drug recommendations and safety warnings. The output is validated server-side and can be stored for traceability and future analysis.

This section presents the main use cases of the intelligent drug recommendation system for common diseases powered by the Gemini API. The scenarios illustrate typical interactions between users and the platform, from account creation to the generation of structured over-the-counter (OTC) medication guidance. Each use case describes how the system captures user inputs, processes clinical context, produces recommendations, and displays safety warnings, maintaining an informative and educational purpose.

Medibot AI was developed to address the need for accessible and structured medication guidance, particularly in contexts where timely access to professional medical consultation may be limited. By integrating Google’s Gemini API, the system leverages large language model capabilities to interpret patient-reported symptoms and clinical context and generate over-the-counter (OTC) medication recommendations accompanied by safety warnings. Unlike traditional clinical decision support tools that rely on fixed rule-based engines or curated drug databases, Medibot AI is designed as a modular web-based system that emphasizes structured prompting, response validation, and traceability of outputs.

A key contribution of this work is the adoption of a structured JSON-based prompting strategy combined with server-side response verification. The system constrains model outputs through strict formatting rules, schema expectations, and validation procedures, reducing the likelihood of malformed or ambiguous recommendations. This is particularly relevant in health-related applications, where unconstrained natural language generation may lead to unclear dosage instructions or unsafe guidance. In addition, the system includes safety-oriented behavior such as suppressing recommendations and returning explanatory reasons when risk conditions are detected (e.g., severe symptoms, red flags, pediatric cases without sufficient context, pregnancy with incomplete information, relevant allergies, or potential drug interactions).

Despite these design measures, the effectiveness of Medibot AI remains dependent on multiple factors. First, the accuracy and safety of recommendations are strongly influenced by the quality and completeness of user-provided data, since missing or incorrect clinical context may limit the system’s ability to detect contraindications or high-risk scenarios. Second, although the Gemini API provides advanced language understanding and reasoning capabilities, its coverage of drug interactions and clinical edge cases may not be fully comprehensive, which can reduce reliability in complex situations. ¹⁹ Therefore, Medibot AI should be interpreted as a guidance-oriented tool, rather than a substitute for professional diagnosis or prescription.

From a data quality perspective, the METRIC framework proposed by ²⁰ highlights essential dimensions such as accuracy, completeness, and representativeness. These dimensions are directly relevant to Medibot AI, since the system operates on patient-reported symptom narratives and contextual inputs that may vary widely across users. Ensuring structured input capture and validation is therefore critical to reduce noise and improve the consistency of generated outputs.

Ethical and professional considerations are central to the responsible deployment of AI in healthcare. The system must preserve patient autonomy, ensure transparency regarding its limitations, and reinforce that outputs do not replace expert evaluation. In this regard, Medibot AI incorporates warnings and safety rules to encourage users to seek professional assistance when symptoms are severe, persistent, or unclear. Additionally, professional responsibility remains essential: healthcare personnel should treat Medibot AI as a complementary decision-support mechanism rather than a replacement for clinical judgment. Adequate training and clear usage protocols are necessary to improve acceptance and prevent misuse. ^{21, 22}

Future work could strengthen the system by improving clinical reliability and expanding its scope. Potential enhancements include integrating additional automated safety layers (e.g., stronger contraindication detection), expanding the coverage of medication scenarios, and incorporating mechanisms that align recommendations with updated scientific evidence. Collaboration with academic institutions and hospitals would be particularly valuable for expert-driven validation, iterative improvement, and formal evaluation of system performance in real-world settings. ²³ Moreover, evaluating the system from the perspective of multiple stakeholders—patients, clinicians, and administrators—could support improvements in usability, integration into healthcare workflows, and long-term sustainability. ²³

Overall, Medibot AI represents a relevant step toward practical AI-assisted medication guidance through a modular architecture and safety-oriented design. While important limitations remain, the system provides a structured and auditable foundation for future research and development aimed at improving access to reliable health information in a responsible manner.

This study presented the design and implementation of Medibot AI, a web-based intelligent system for informational over-the-counter (OTC) medication guidance using the Gemini API. The system demonstrates that large language models can be integrated into a modular architecture capable of transforming user-reported symptoms and clinical context into structured prompts, producing schema-constrained JSON outputs, and delivering recommendations accompanied by safety warnings. The incorporation of server-side validation, persistence mechanisms, and traceability supports robustness and reproducibility of the generated results.

Despite these contributions, the system is subject to limitations inherent to LLM-based solutions. Recommendation quality depends on the accuracy and completeness of user-provided inputs and on the model’s ability to generate consistent structured responses. Although the proposed validation and safety rules reduce the risk of malformed or unsafe outputs, Medibot AI is not intended to provide diagnosis or replace professional medical advice. Therefore, its outputs should be interpreted as educational guidance, and users should consult healthcare professionals when symptoms are severe, persistent, or unclear.

Future work may focus on strengthening clinical reliability and scalability. This includes expanding the medication knowledge base, improving safety coverage for contraindications and drug interactions, and integrating retrieval-based strategies (e.g., RAG) to ground recommendations in verified pharmacological sources. Additional improvements could include multilingual support, more refined risk stratification, and systematic evaluation protocols involving expert review and real-world usability studies.

Overall, Medibot AI provides a practical contribution toward responsible AI-assisted medication guidance by combining structured prompting, validation, and safety-oriented design. These elements support the development of transparent and auditable systems that can enhance access to reliable health information while maintaining ethical and clinical caution.

This article describes the development of a software tool and does not report results from a clinical study or interventions involving patients. The system is intended for informational and educational purposes only and does not provide medical diagnosis, clinical decision-making, or professional medical advice. Medibot AI is not a medical device. Users voluntarily enter information to obtain guidance related to over-the-counter medications. The interface includes visible safety warnings encouraging consultation with a healthcare professional, particularly when symptoms are severe, persistent, or unclear.

Software available from: https://medibot-ai-prod.vercel.app/

Source code available from: https://github.com/PIEROLS15/medibot-ai

Archived source code at time of publication: https://doi.org/10.5281/zenodo.17714829 . ²⁵

License:* MIT License — https://opensource.org/license/MIT

The system stores user account information (e.g., name and email) and consultation history (e.g., symptoms and related form fields) in a PostgreSQL database to support traceability and reproducibility of outputs. Access requires authentication. API credentials are protected using environment variables, and input validation is enforced using Zod to reduce the risk of malformed data submissions.

No real patient data were used or publicly shared. The publicly archived dataset consists exclusively of synthetic scenarios intended for reproducibility and validation purposes.