During the data collection phase, we crawled and cleaned data from the NAVER Knowledge iN platform, as depicted in part A of Figure 1. The Knowledge iN platform, which allows a wide range of users to post questions and answers, is very popular in Korea [14]. We collected question–answer pairs from the official profile of the Youth Cyber Counseling Center, managed by the Korea Youth Counseling and Welfare Institute, a quasi-governmental agency under the Ministry of Gender Equality and Family and an official NAVER partner. All publicly available responses from this verified profile, along with their corresponding questions, were included. Python-based Selenium was used to extract the content of the answers and the text of the corresponding questions [15]. The extracted data spans a period of 10 years, from 2014 to 2024, with each post containing a youth’s question and an expert’s answer.

During the crawling stage, entries that mentioned age or school grade indicative of adults or children were excluded to increase the likelihood that the questions were authored by adolescents. We also ensured that each entry contained both a question and an answer to remove cases with missing values.

The collected data were preprocessed by taking several steps to eliminate unnecessary content. First, we removed 2,598 cases in which the question contained 15 words or less, as such entries provided few qualifying examples for analysis. Second, we eliminated 1,995 cases where the question or answer was duplicated across different entries. Third, we discarded data that did not express negative sentiment by applying an automated, keyword-based classification using the emotion lexicons from Dheeraj and Ramakrishnudu [16], rather than manual annotation. Among the five original categories, “addiction” was excluded for cultural reasons, and only depression, anxiety, stress, and obsessive-compulsive disorder (OCD) were used for filtering. A detailed list of keywords is provided in Supplement A. Finally, we removed data that did not reveal adolescent distress. To filter out data lacking clear expressions of worry, we categorized the questions into general and detailed causes based on the study of Hellstrom et al. [17]. This classification was appropriate for our dataset, as the primary categories—such as academic stress, family conflict, and peer relationships—are also recognized as key mental health concerns among Korean adolescents in recent studies [4,18]. General causes include major categories of worry (e.g., love, friendship, health), while detailed causes enumerate specific concerns under each general category. A detailed list of keywords is provided in Supplement B. After screening, 705 questions that did not reveal a cause were removed. The final dataset consists of 3,874 questions.

In the answer generation phase, the cleaned data was used to improve the LLM’s responses through the RAG method, as illustrated in part B of Figure 1. RAG is a technique that enables an LLM to search for external information to formulate answers, thereby preventing the generation of illusory or overly generic content. In this study, we leveraged this feature of RAG to improve answer quality by retrieving data that are well-aligned with each specific problem and then generating more reliable responses [13].

During answer generation, the “all-MiniLM-L12-v2” model from Sentence Transformers (using a 384-dimensional embedding) was employed to convert the collected data into vector form for storage [19]. When an answer was requested, the input question was encoded and the top-k most similar vectors were retrieved using FAISS with L2 distance, without the application of a fixed similarity threshold [20]. Subsequently, the LLM model was used to enhance the generated answer by referencing the cases with the highest similarity. In this study, the top three cases with the greatest similarity were presented to the GPT-4o mini model, which then generated an answer for the given question [21]. The GPT-4o mini generation parameters are detailed in Supplement C.

To evaluate the performance of the generated answers, we constructed an evaluation dataset as shown in part C of Figure 1. This dataset was assembled under identical conditions, except for the use of RAG. It comprises 3,874 final data points, evenly distributed across the four negative sentiment categories, totaling 781 points. Each data point includes a question and two sets of answers: one generated by the RAG model and one generated without RAG.

An automated evaluation was performed using the Gemini model to validate the generated answers [22]. Four evaluation metrics were selected—relevance, comprehensibility, empathy, and practicality—and specific prompts were organized for each metric. For instance, the Relevance metric assesses how closely the answer aligns with the question. Each metric was used to compare the two generated answers and select the superior one. Detailed definitions of each metric, along with supporting literature, are provided in Table 1, and the prompts are available in Supplement D [23,24]. The Gemini 1.5 Flash evaluation parameters are detailed in Supplement E.

This study comprises 3,874 cases derived from youth Q&A data collected between 2014 and 2024, categorized by four major negative emotion keywords: depression, anxiety, stress, and OCD. The distribution of the cases by year and the percentage attributed to each emotion are presented in Table 2.

Figure 2 shows a heatmap that visualizes the frequency with which different negative emotions co-occur. Depression and stress frequently appeared together, with 615 instances, while depression and anxiety were paired 376 times, indicating a strong association between these emotions. Conversely, OCD was rarely mentioned alongside other major emotions. Overall, depression was the emotion most commonly seen in combination with other feelings, suggesting that it may serve as a central element in complex emotional experiences. In contrast, OCD tended to appear more independently. These findings indicate that some emotions frequently co-occur rather than appearing in isolation and that the combinations of emotions in the dataset are quite varied.

To analyze the main causes of distress, the study visualized keyword frequency and associations. Text normalization was applied to reduce redundant expressions such as “because of my weight” and “because of my height” to their essential elements, “weight” and “height,” respectively.

Figure 3 displays a word cloud that visually represents the detailed causes of adolescent distress. The size of each keyword reflects its frequency, highlighting the primary areas of worry among adolescents. Dominant keywords such as “friend,” “cram school,” “love,” and “test” suggest that stress related to academics, relationships, and friendships is common. In addition, family-related keywords like “mother,” “father,” and “parents” indicate that conflicts at home and issues with parental relationships are major stressors.

Figure 4 presents a network graph that illustrates the relationships between general cause and detailed cause in adolescents’ worries. The node colors represent the major categories of distress, and red edges and nodes emphasize the top-10 keyword pairs—those that co-occur most frequently in the data.

Table 3 lists the top-10 keyword pairs, organized by frequency. The most common pairs are “Parents - Friend” (348 times), “Father - Mother” (325 times), and “Mother - Friend” (308 times), suggesting that adolescents often discuss conflicts or problems with friends in the context of parental relationships, especially with mothers. Conversely, the data also indicate that parental advice or intervention can influence friendships. This pattern, characterized by concurrent references to parents and friends in distress narratives, suggests that adolescents view home and friendship matters as closely intertwined. Additionally, pairs such as “Love - Friend” (258 times), “Friend - Cram School” (258 times), and “Test - Friend” (195 times) reflect the complex interplay between academic pressures and friendship issues.

Tables 4 and 5 summarize the evaluation results comparing answers generated by the RAG and non-RAG approaches. Table 4 details the quantitative evaluation outcomes, indicating that RAG-based answers performed better across all metrics.

Table 5 provides a qualitative comparison of the answer quality produced by each method. The evaluation revealed that RAG-generated responses surpassed non-RAG answers on all metrics. For example, RAG-based answers addressed the user’s concerns with specificity and offered clear guidance for resolution. One response—“I lose confidence due to comparison with others in dance school…”—was highly rated because it pinpointed the core issue and suggested a concrete solution. In contrast, non-RAG responses tended to offer generic encouragement, such as “Everyone feels awkward at first,” and lacked personalized advice.

RAG-based answers also scored well on the empathy metric. A response containing the phrase “It’s not easy to adapt to a new environment” comforted the user by acknowledging their feelings in a natural, understanding manner. Non-RAG responses, while supportive, did not effectively integrate empathy with problem solving. RAG-based responses distinguished themselves by balancing empathetic understanding with practical solutions, which helped users feel both understood and supported.

The comprehensibility metric also favored RAG-based answers. These responses organized information logically and employed structured, step-by-step language (e.g., “first, then…”) that made it easier for users to follow the advice. In contrast, non-RAG responses typically presented disjointed sentence lists that lacked coherent flow.

Finally, the practicality of RAG-based answers was evident in their provision of specific, actionable instructions. For example, advice such as “Set a small goal to master one new move in every class, or ask a question at least once during class” offered concrete, practical steps for improvement. This approach helps users build confidence incrementally. In contrast, another RAG response—“Learning to dance and gaining confidence is a process that takes time. It’s important to stick with it and not judge your results based on one or two days of experience”—offered a more abstract encouragement without a concrete plan, making it less actionable.

In addition to the example presented in Table 5, a second case involving self-esteem and appearance-related concerns— common themes in our dataset—is provided in Supplement E. This additional example was evaluated using the same four metrics and further reinforces the consistent strengths of RAG-based responses across a range of topics.

The analysis revealed that adolescents’ worries are influenced by a range of factors, primarily family relationships, peer relationships, and academic stress, all of which serve as major sources of emotional strain. In particular, conflicts within the family, especially between parents, significantly shape adolescents’ emotional experiences, while issues in peer relationships—such as alienation, conflict, and bullying— undermine their psychological well-being. Academic stress is largely driven by the pressures of maintaining competitive grades and preparing for exams, pressures that many adolescents attempt to mitigate by sharing their experiences with friends and seeking emotional support. These findings suggest that youth response models must extend beyond addressing simple issues to recognizing and addressing complex problems in an integrated manner.

RAG-based responses outperformed non-RAG responses across all four evaluation metrics. They excelled on the Practicality metric by providing concrete, step-by-step solutions, and they achieved high scores on the Empathy metric by combining understanding with actionable advice. In contrast, non-RAG responses tended to be more generic and abstract. These results indicate that integrating search augmentation can significantly increase both the specificity and overall effectiveness of personalized answer generation, particularly in tackling the complex concerns faced by young people.

Although this study effectively leveraged youth Q&A data to visualize and analyze various distress factors and their associations, several limitations remain.

First, the reliance on keyword-based categorization presents challenges. The same keyword can have different meanings in various contexts, necessitating careful interpretation of the analysis. For example, the term “friend” might indicate conflict in one context while simply referring to a person in another. Furthermore, overlapping categorizations can lead to inflated or duplicated frequency counts when a single case is assigned to multiple categories.

Second, our data were sourced exclusively from the NAVER Knowledge iN platform rather than from multiple external services. This may introduce biases related to age, gender, and socioeconomic background that do not necessarily reflect the broader population of Korean adolescents.

Third, limitations exist in the application of LLM models. While this study utilized the GPT-4.0 model combined with RAG for answer generation, comparisons with open-source models such as LLaMA or with other models like Gemini were not conducted [22,25].

Fourth, the method of data partitioning presents another limitation. The dataset was divided into reference and test sets in an 8:2 ratio, which might affect the model’s generalizability when evaluated against external data.

This study confirmed the effectiveness of a customized answer generation model that incorporates RAG technology based on Korean adolescent Q&A data. As demonstrated by the case analysis in Table 5, RAG-based answers more accurately capture users’ specific concerns while providing both relevant information and actionable advice. These results underscore the importance of developing an answer model that can effectively address the diverse challenges encountered by young people in a digital environment [26].

Future research should focus on constructing a dataset that more deeply reflects the cultural context and social environment of Korean adolescents and on further improving the algorithm’s answer quality and intuitiveness by incorporating a wider range of user feedback. Additionally, it is important to design an answer model capable of supporting a richer problem-solving experience by enhancing its ability to handle multi-turn conversations and complex scenarios. Beyond merely dispensing information, a customized answer service that holistically considers each user’s unique background and challenges will be instrumental in addressing the multi-faceted problem-solving needs of adolescents [27].