Paper deep dive
TRACE: A Conversational Framework for Sustainable Tourism Recommendation with Agentic Counterfactual Explanations
Ashmi Banerjee, Adithi Satish, Wolfgang Wörndl, Yashar Deldjoo
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 95%
Last extracted: 4/18/2026, 1:24:24 AM
Summary
TRACE is a multi-agent, LLM-based conversational recommender system designed to promote sustainable tourism. It utilizes an orchestrator-worker architecture to elicit user preferences, construct personas, and provide recommendations that balance relevance with environmental impact. The system employs agentic counterfactual explanations and clarifying questions to nudge users toward greener travel choices without coercion, maintaining high user trust and recommendation quality.
Entities (5)
Relation Signals (3)
TRACE → uses → Google Agent Development Kit
confidence 100% · TRACE is implemented as a modular, state-aware pipeline using Google’s Agent Development Kit (ADK)
Clarifying Question Agent → partof → TRACE
confidence 95% · The proposed system adopts a modular, microservices-based architecture... consisting of a sequential pipeline of specialized LLM agents
Explanation Generation Agent → partof → TRACE
confidence 95% · The Explanation Generation Agent serves as the final decision-making component.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Traditional conversational travel recommender systems primarily optimize for user relevance and convenience, often reinforcing popular, overcrowded destinations and carbon-intensive travel choices. To address this, we present TRACE (Tourism Recommendation with Agentic Counterfactual Explanations), a multi-agent, LLM-based framework that promotes sustainable tourism through interactive nudging. TRACE uses a modular orchestrator-worker architecture where specialized agents elicit latent sustainability preferences, construct structured user personas, and generate recommendations that balance relevance with environmental impact. A key innovation lies in its use of agentic counterfactual explanations and LLM-driven clarifying questions, which together surface greener alternatives and refine understanding of intent, fostering user reflection without coercion. User studies and semantic alignment analyses demonstrate that TRACE effectively supports sustainable decision-making while preserving recommendation quality and interactive responsiveness. TRACE is implemented on Google's Agent Development Kit, with full code, Docker setup, prompts, and a publicly available demo video to ensure reproducibility. A project summary, including all resources, prompts, and demo access, is available at this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2604.14223v1
- Canonical: https://arxiv.org/abs/2604.14223v1
Trouble viewing inline? Open PDF directly →
Full Text
33,701 characters extracted from source content.
Expand or collapse full text
TRACE: A Conversational Framework for Sustainable Tourism Recommendation with Agentic Counterfactual Explanations Ashmi Banerjee ashmi.banerjee@tum.de Technical University of Munich Munich, Germany Adithi Satish adithi.satish@tum.de Technical University of Munich Munich, Germany Wolfgang Wörndl woerndl@in.tum.de Technical University of Munich Munich, Germany Yashar Deldjoo yashar.deldjoo@poliba.it Polytechnic University of Bari Bari, Italy Abstract Traditional conversational travel recommender systems primar- ily optimize for user relevance and convenience, often reinforc- ing popular, overcrowded destinations and carbon-intensive travel choices. To address this, we present TRACE (Tourism Recommen- dation with Agentic Counterfactual Explanations), a multi-agent, LLM-based framework that promotes sustainable tourism through interactive nudging. TRACE uses a modular orchestrator-worker architecture where specialized agents elicit latent sustainability preferences, construct structured user personas, and generate rec- ommendations that balance relevance with environmental impact. A key innovation lies in its use of agentic counterfactual explana- tions and LLM-driven clarifying questions, which together surface greener alternatives and refine understanding of intent, fostering user reflection without coercion. User studies and semantic align- ment analyses demonstrate that TRACE effectively supports sus- tainable decision-making while preserving recommendation quality and interactive responsiveness. TRACE is implemented on Google’s Agent Development Kit, with full code, Docker setup, prompts, and a publicly available demo video to ensure reproducibility. A project summary, including all resources, prompts, and demo access, is available at https://ashmibanerjee.github.io/trace-chatbot. CCS Concepts • Information systems→Information retrieval;• Computing methodologies→ Artificial intelligence. Keywords Conversational Recommender Systems, LLMs, Multi-Agent Sys- tems, Sustainable Tourism, Counterfactual Explanations ACM Reference Format: Ashmi Banerjee, Adithi Satish, Wolfgang Wörndl, and Yashar Deldjoo. 2026. TRACE: A Conversational Framework for Sustainable Tourism Recommen- dation with Agentic Counterfactual Explanations. In Proceedings of the 49th This work is licensed under a Creative Commons Attribution-NonCommercial- NoDerivatives 4.0 International License. SIGIR ’26, Melbourne, VIC, Australia © 2026 Copyright held by the owner/author(s). ACM ISBN 979-8-4007-2599-9/2026/07 https://doi.org/10.1145/3805712.3808370 International ACM SIGIR Conference on Research and Development in Infor- mation Retrieval (SIGIR ’26), July 20–24, 2026, Melbourne, VIC, Australia. ACM, New York, NY, USA, 6 pages. https://doi.org/10.1145/3805712.3808370 1 Introduction Conversational agents are increasingly used to support complex decision-making tasks such as travel planning. Recent advances in LLMs have enabled chat-based travel assistants that can flexibly respond to natural language queries and generate personalized rec- ommendations [19,31,33]. However, most existing Conversational Recommender Systems (CRS) for travel primarily optimize for user relevance and convenience, often reinforcing popular, overcrowded destinations and carbon-intensive travel choices. As a result, sus- tainability considerations, such as environmental impact, conges- tion, and seasonality, are typically treated as secondary constraints, if considered at all [2, 12, 35]. In this demo paper, we present TRACE (Tourism Recommen- dation with Agentic Counterfactual Explanations), an LLM-agent- based conversational recommender system explicitly designed to promote sustainable tourism practices. Unlike traditional chatbots or chat-based travel assistants, TRACE incorporates sustainability objectives from the ground up. The system prioritizes recommen- dations such as less-visited or emerging destinations over over- crowded hotspots, and favors lower-impact transportation options (e.g., train travel over flights) whenever feasible. Crucially, these sustainability signals are not hard-coded defaults but are inferred through interaction. The primary goal of this work is not to build a state-of-the-art recommender system in terms of predictive accuracy or ranking performance. Instead, we aim to explore the design space and re- search opportunities enabled by multi-agent LLM architectures for sustainable conversational recommendation. In particular, we focus on how such systems can be structured to nudge users toward more sustainable travel choices without violating their explicit prefer- ences or degrading user trust. TRACE achieves this through a modular, agentic design. Dedi- cated agents generate targeted clarifying questions to elicit latent sustainability preferences, construct structured user personas, and produce recommendations that jointly optimize for relevance and sustainability signals. A key contribution of the system is its expla- nation agent, which generates persuasive justifications highlighting the sustainability benefits of recommended options. When users arXiv:2604.14223v1 [cs.IR] 14 Apr 2026 SIGIR ’26, July 20–24, 2026, Melbourne, VIC, AustraliaAshmi Banerjee, Adithi Satish, Wolfgang Wörndl, and Yashar Deldjoo do not explicitly express sustainability preferences, the system em- ploys counterfactual explanations [20] to gently expose greener alternatives, enabling informed reflection rather than coercive in- tervention. We demonstrate TRACE as an interactive system that showcases how agentic LLM pipelines can operationalize responsible nudg- ing in tourism recommendations. By releasing this demo, we aim to stimulate discussion on the role of conversational interfaces, explanations, and counterfactual reasoning in shaping more sus- tainable user behavior and to encourage further research beyond accuracy-centric evaluation paradigms in recommender systems 1 . Contributions. This paper introduces TRACE, a multi-agent, LLM- based conversational recommender system for sustainable tourism: •Modular, multi-agent framework: Specialized agents han- dle user modeling, clarifying questions, recommendations, and explanations, integrating digital nudging to promote greener choices. •Open-source release and demo: The full code and frame- work are publicly available, accompanied by a demo video to facilitate reproducibility and further research 2 . •Empirical evaluation: User studies and semantic metrics show TRACE balances relevance, and sustainability nudging while maintaining interactive response times. Related Work. Conversational recommender systems have been widely studied for interactive preference elicitation and natural language decision support. Recent LLM-powered, multi-agent ap- proaches leverage specialized agents for user modeling, planning, and explanation, enabling collaborative recommendation and flexi- ble reasoning [22,25,27]. However, these systems primarily focus on accuracy and engagement, leaving sustainability and behavior change as secondary concerns [12,35]. Although recent work [6] has explored multi-agent frameworks to balance relevance and sustainability in tourism recommendations, these approaches do not explicitly incorporate conversational interactions or leverage explanations and counterfactuals to nudge user behavior, and they remain primarily research prototypes with limited production readi- ness. Explainable and counterfactual recommendation research fur- ther explores transparency and bias mitigation, often via minimally contrastive explanations [8,20,32,34,36]. Relatedly, clarifying ques- tions have been investigated as a way to uncover hidden or incom- plete user preferences [9,28–30], with taxonomies spanning vague, spatial, and temporal dimensions [37] and LLM-based dialogue updates supporting personalization [24]. In parallel, sustainability- aware recommenders apply digital nudging and eco-objectives to guide greener choices [3–5,21,26], though most remain single- agent, non-conversational, rely on static sustainability encodings, and often do not leverage LLMs. In contrast, TRACE integrates these directions in a multi-agent, LLM-based conversational frame- work, where agents infer preferences, build structured personas, and explain greener choices as an evolving dialogue objective. 1 Demo video: https://youtu.be/BdtEFSp42fw 2 https://ashmibanerjee.github.io/trace-chatbot Conversational Interface User input ... q ∈ Q Central Orchestrator Clarifying Questions Agent (A CQ ) Intent Classifier Agent (A IC ) Explanation Generation Agent (A EG ) Rec-LLM Agent (A Rec ) ADK Multi-Agent (LLM) Backend Agentic Reasoning Layer (Backend) Orchestration Layer (Middleware) User Interaction Layer (Frontend) (R 0 , R 1 ) State Management & Persistent Storage Figure 1: System Architecture of the CRS Chatbot. This dia- gram illustrates the Orchestrator-Worker paradigm, detail- ing the flow between the User Interaction layer (Frontend), Orchestration layer (Middleware), and the Agentic Reason- ing layer (Backend). 2 TRACE Framework: System Design The proposed system adopts a modular, microservices-based archi- tecture for delivering sustainable travel recommendations using a Multi-Agent LLM framework. The design follows an Orchestrator- Worker paradigm, where a central middleware component coordi- nates user interactions and delegates reasoning tasks to specialized agents. The architecture is organized into three conceptual layers: • User Interaction Layer (Frontend), responsible for con- versational interaction and feedback collection. •Orchestration Layer (Middleware), which manages ses- sion state, control flow, and agent execution. •Agentic Reasoning Layer (Backend), consisting of a se- quential pipeline of specialized LLM agents implemented using Google’s Agent Development Kit (ADK) over vertexAI. Throughout this paper, we define context as the information accumulated from a user’s initial query and responses to clarifying questions, which collectively inform the recommendation process by capturing inferred preferences and sustainability considerations. This layered design promotes modularity, scalability, and clear sepa- ration of concerns between interaction handling, control logic, and reasoning components. Figure 1 illustrates the overall system ar- chitecture, highlighting the interactions between the user interface, orchestrator, and various LLM agents. 2.1 Formalization of the Multi-Agent CRS We formalize the TRACE Conversational Recommender System (CRS) as a sequential pipeline of transformations over the state space of a user session. Let푞 ∈ Qdenote the initial natural language query. The system proceeds through the following stages. 2.1.1 Elicitation and Persona Modeling. The Clarifying Question Agent퐴 CQ generates a set of targeted questions퐶=푐 1 ,푐 2 , . . .,푐 푛 to elicit latent user preferences related to sustainable travel: 퐶= 퐴 CQ (푞).(1) TRACE: Tourism Recommendation with Agentic Counterfactual ExplanationsSIGIR ’26, July 20–24, 2026, Melbourne, VIC, Australia The primary goal of the Clarifying Question Agent is to elicit the user’s sustainability preferences indirectly and their willingness to compromise, for example, by asking whether they would be willing to travel to a lesser-known place rather than a popular (and crowded) destination, thereby improving intent detection. However, if the user’s initial query is vague ("I want to travel in Europe"), then퐴 CQ is also instructed to ask the user about general travel aspects such as their budget and interests. The user’s initial query is provided as input to퐴 CQ , which generates up to 5 clarifying questions, which are then shown to the user one at a time. Given the user responsesΓto these questions, the Intent Clas- sification Agent퐴 IC maps the interaction into a structured User Travel PersonaUand a vector representing the user’s Willingness to Compromise (WTC) across sustainability dimensions such as emissions, congestion, and seasonality: (U, WTC)= 퐴 IC (Γ,푞).(2) The generated User Travel Persona and Willingness to Compromise are provided to the recommender to obtain personalized, context- aware recommendations. 2.1.2 Recommendation Generation. The system constructs two candidate recommendation sets. The baseline set푅 0 is instructed to prioritize relevance solely to the original query푞. In contrast, the sustainable set푅 1 incorporates the inferred user persona and sustainability preferences from퐴 IC . Both sets are generated by the Recommender Agent퐴 Rec , a Rec-LLM using few-shot prompting. In particular,푅 1 is produced by prioritizing sustainability-related sig- nals푆derived from the clarification responsesΓ, while preserving alignment with the inferred personaU: 푅 1 = 퐴 Rec (U,Γ,푆).(3) 2.1.3 Persuasive Explanation and Decision Logic. The Explanation Generation Agent퐴 EG serves as the final decision-making compo- nent. It outputs a selected recommendation푟 ∗ , a persuasive expla- nation퐸, and a counterfactual alternative푟 alt , explicitly conditioned on the inferred willingness to compromise: (푟 ∗ ,퐸,푟 alt )= 퐴 EG (푅 0 ,푅 1 ,U, WTC).(4) The agent dynamically adopts one of two rhetorical strategies: Direct Alignment. IfWTCindicates openness to sustainability trade-offs, the agent selects푟 ∗ ∈ 푅 1 and generates an explanation emphasizing the sustainability improvement Δ푆= 푀(푅 1 )− 푀(푅 0 ), where푀(·)denotes a vector of sustainability metrics (e.g., CO 2 emissions, visitor pressure, walkability). Counterfactual Nudging. IfWTCindicates resistance to sustain- ability trade-offs, the agent selects the baseline recommendation 푟 ∗ ∈ 푅 0 to preserve user trust, while generating a counterfactual explanation for an alternative 푟 alt ∈ 푅 1 . The explanation is framed conditionally, for example: “Had you expressed interest in lower en- vironmental impact,푟 alt would have been recommended because ...” This strategy enables implicit nudging toward sustainable options without violating the user’s explicitly stated constraints. 2.2 Implementation TRACE is implemented as a modular, state-aware pipeline using Google’s Agent Development Kit (ADK) [16] on Vertex AI [17], with thegemini-2.5-flashmodel handling all agentic reason- ing [18]. The backend uses an Orchestrator-Worker pattern man- aged via FastAPI [13], with Google Firestore [14] maintaining persistent session state (user queries, persona vectors, and푊푇퐶). The user interface, built with Chainlit [10], supports a seamless Clarifying Question loop and the rendering of recommendations. The stack is containerized with Docker and deployed on Google Cloud Run [15] for scalable, serverless execution of asynchronous multi-agent workflows. User Travel Profile: A couple seeking a romantic European getaway, drawn to lakeside or seaside destinations with great food and nightlife, who value authentic, less-known, less crowded experiences and prefer mid-range to luxury travel. Your Recommendation: Valencia (sustainable ) You’re looking for nightlife and food experiences by the sea, with a less crowded and unique vibe. Valencia delivers vibrant cuisine, lively nightlife, and a Mediterranean setting—offering a unique, less-known alternative to crowded cities like Barcelona. Alternative Option: Barcelona While Barcelona also offers excellent nightlife, food, and a seaside location, it is a highly popular destination. This could lead to a more crowded experience, which you indicated you were willing to compromise on for a more unique and less-known destination. Which recommendation did you like more? Me and my boyfriend are looking for somewhere to spend vacation in Europe, where there is either a lake or a seaside, where would you recommend? Query (q) midrange to luxury What's your approximate budget for this trip (e.g., budget-friendly, mid-range, luxury)? Beyond the lake or seaside, what kind of activities or atmosphere are you hoping for? nightlife and food experience Clarifying Questions - A CQ (q) Would you try a lesser-known lakeside or seaside spot for a more unique, less-crowded experience? yes of course ✨Analyzing your preferences... This may take a moment. Intent Classifier - A IC r* ∈ R 1 r alt ∈ R 0 Expl. Gen. A EG Expl. Gen. A EG Valencia Feedback continues ... A single session Figure 2: An example TRACE session. The session starts with the user’s initial query, followed by clarifying questions. It concludes with a generated travel profile and a set of recom- mendations, each accompanied by explanations. The session also collects session-specific feedback, including which rec- ommendations were preferred, how helpful the explanations were, how effective the clarifying questions were, and an op- tion to provide qualitative feedback. Users can enter free-text queries or select predefined travel sce- narios that cover a range of preferences and sustainability con- siderations (e.g., eco-friendly stays, off-the-beaten-path trips). We use sample queries from SynthTRIPS [7], which provide synthetic travel queries and preferences, to offer users inspiration for their initial inputs. To prevent misuse, the Clarifying Question Agent (퐴 CQ ) enforces guardrails with few-shot instructions, restricting recommendations for single European city trips. Queries outside this scope (for exam- ple, "Recommend some movies to watch this weekend") are flagged as invalid, and users are prompted to stay within city trip recom- mendations. 2.3 User Journey In this paper, a session denotes a complete interaction between a user and the system, beginning with the user’s initial query and ending with their responses to the feedback questions. Each session proceeds as follows. Given their query, the user first responds to a set of clarifying questions to construct a travel per- sona. The system then generates two candidate recommendations: SIGIR ’26, July 20–24, 2026, Melbourne, VIC, AustraliaAshmi Banerjee, Adithi Satish, Wolfgang Wörndl, and Yashar Deldjoo a baseline recommendation푅 0 and a context-aware recommenda- tion푅 1 . Based on the user’s willingness to compromise (WTC) on sustainability, the Explanation Generation Agent determines which candidate is presented as the primary recommendation,푟 ∗ , and which is shown as the alternative,푟 alt . When the user indicates no willingness to compromise on sustainability, the agent may assign the baseline recommendation (푅 0 ) as the primary option. Figure 2 illustrates an example session in which the user ex- presses a preference for a seaside holiday. In this instance, openness to exploring lesser-known destinations to avoid crowds results in Valencia (푅 1 ) being presented as the primary recommendation (푟 ∗ ), with Barcelona (푅 0 ) shown as the alternative (푟 alt ). Had the user instead preferred popular destinations without such flexibility, the roles of the two recommendations would have been reversed. 3 Evaluation and User Insights We evaluate TRACE using a user study and quantitative alignment analysis to assess its ability to promote sustainable travel choices and support coherent agentic reasoning. 3.1 User Study Design We recruited 24 participants (58.3% male, 41.7% female) via social media, personal networks, and university mailing lists. Most partici- pants were aged 18–34 (79%), and 54% reported frequent chatbot use. Participants interacted with TRACE using 1–2 travel queries and were encouraged to include challenging or out-of-scope requests (e.g., non-European destinations) to test robustness. After the interaction, participants rated the system on a 5-point Likert scale (1: Not at all, 5: Extremely) [23] across three dimensions: quality of clarifying questions, persuasiveness of explanations, and extent of choice reconsideration. After filtering incomplete and out- of-scope responses, we analyzed 푁= 107 valid conversations. The evaluation examines whether TRACE can nudge users to- ward sustainable choices without being coercive, while maintaining alignment across agents and with user intent. We formalize these objectives through the following research questions (RQs): (1)User Feedback and Interaction Analysis (RQ1): How ef- fectively does the system promote sustainable choices while preserving user trust and engagement? (2)System Alignment and Semantic Metrics (RQ2): To what extent does the system maintain alignment across its internal agents and with user intent? (3)System Latency (RQ3): Does TRACE achieve response times suitable for real-time interaction despite the computational overhead of multiple agents? 3.2 RQ1: User Feedback and Interaction Analysis RQ1 investigates whether TRACE promotes sustainable choices while preserving user trust and engagement. Results from the user study (Figure 3) show high acceptance of the system’s conversa- tional interaction. When presented with a primary and alternative recommendation, 79.1% of users selected the primary option (푟 ∗ ), which in 75.5% of sessions corresponded to the context-aware (typi- cally more sustainable, i.e.,푅 1 ) recommendation. In sessions where the baseline (푅 0 ) was presented as the primary option (푟 ∗ ), 16.7% of users selected the context-aware alternative (푅 1 ), indicating a modest nudging effect. User feedback further supports the effectiveness of the interac- tion design: 55.2% of users rated the clarifying questions and 65.4% rated the explanations as Very Well or Extremely Well. Moreover, around 60% of users reported some degree of reconsideration of their initial choice, suggesting that TRACE encourages reflection while preserving user agency. Overall, these findings indicate that TRACE can support sustainable recommendations without com- promising user engagement or trust. 020406080100 Percentage of Responses (%) Clarifying Questions Quality Explanation Quality Choice Reconsideration Level 15.5 16.4 38.9 8.6 5.5 16.7 20.7 12.7 20.4 48.3 52.7 22.2 6.9 12.7 Response 1 - Not at all 2 - Slightly 3 - Moderately 4 - Very well / Strongly 5 - Extremely Figure 3: Combined feedback distribution across rating cate- gories for Clarifying Question Quality, Explanation Quality, and Choice Reconsideration Level. 3.3 RQ2: System Alignment and Semantic Metrics RQ2 examines whether TRACE maintains alignment between the conversation context, the structured user model, and the explana- tions generated by the agentic pipeline. Using semantic similarity metrics computed with theall-MiniLM-L6-v2model, we measure coherence across these components. The explanations generated by퐴 퐸퐺 show high similarity to the full conversation (User Query 푞+ Clarifying Questions), with a mean score of 0.7033, indicating strong contextual grounding. The output of the Intent Classifier (User Persona, Travel Intent, and WTC) also aligns closely with the conversation (0.7883), confirming accurate intent capture. Finally, the high similarity between the Intent Classifier output and the generated explanations (0.7437) demonstrates that퐴 퐸퐺 consistently conditions its reasoning on the inferred user model and sustain- ability dimensions. Overall, these results indicate strong alignment across TRACE’s internal agents and with user intent. 3.4 RQ3: System Latency We assess whether TRACE can operate within latency bounds suit- able for interactive use despite its multi-agent design. Following the clarifying question phase, TRACE requires an average of 23 seconds to generate user profiles, recommendations, and explana- tions, with a maximum latency of 38 seconds. These results show that TRACE’s modular architecture does not introduce prohibitive overhead, supporting the practical feasibility of multi-agent con- versational recommendation. Moreover, its latency is competitive with that of contemporary multi-agent frameworks [11], which often exceed 40 seconds. Overall, TRACE effectively balances the TRACE: Tourism Recommendation with Agentic Counterfactual ExplanationsSIGIR ’26, July 20–24, 2026, Melbourne, VIC, Australia computational cost of agentic reasoning with the need for accurate and personalized tourism recommendations. 4 Conclusion We presented TRACE, a multi-agent LLM-based conversational recommender system for sustainable tourism, combining user mod- eling, clarifying questions, recommendations, and explanations to nudge users toward greener choices while preserving user trust. User studies and semantic metrics show TRACE balances relevance, trust, and sustainability, nudging with interactive response times. Due to a Chainlit vulnerability discovered on 21 January, 2026 [1], the live app is currently offline, but the code, Docker setup, prompts, and a demo video are publicly available for reproducibility. The system ran successfully from January 16–31, 2026, and can be set up locally using our resources. Currently limited to single European city travel, TRACE will be extended in future work to multi-city, multi-day itineraries. Overall, it demonstrates the feasibility of combining multi-agent reasoning, LLMs, and sustainability-aware conversational recommendation in a practical, reproducible framework. At the same time, TRACE raises an important sustainability paradox. By consistently nudging users toward less popular “hidden gems,” the system may inad- vertently create new hotspots, shifting rather than reducing over- tourism. Addressing this requires dynamic, adaptive recommenda- tion strategies that account not only for individual preferences but also for destination-level impacts. Future work should therefore incorporate real-time signals, such as destination capacity, envi- ronmental indicators, and user feedback, to continuously refine recommendations and mitigate such unintended consequences. GenAI Usage Disclosure We used ChatGPT, Claude, and Gemini for code suggestions, and Grammarly for language refinement; all outputs were critically reviewed, and we take full responsibility for the final version. Acknowledgments We thank the Google AI/ML Developer Programs team for support- ing us with Google Cloud Credits. References [1]Ionut Arghire. 2026.Chainlit Vulnerabilities May Leak Sensitive Infor- mation.https://w.securityweek.com/chainlit-vulnerabilities-may-leak- sensitive-information/. Accessed 2026-02. [2]Ashmi Banerjee. 2023. Fairness and sustainability in multistakeholder tourism recommender systems. In Proceedings of the 31st ACM Conference on User Modeling, Adaptation and Personalization. 274–279. [3] Ashmi Banerjee, Melih Mert Aksoy, and Wolfgang Wörndl. 2025. SmartSustain Recommender System: Navigating Sustainability Trade-offs in Personalized City Trip Planning. arXiv preprint arXiv:2510.17355 (2025). [4]Ashmi Banerjee, Tunar Mahmudov, Emil Adler, Fitri Nur Aisyah, and Wolfgang Wörndl. 2025. Modeling sustainable city trips: integrating CO 2 e emissions, popularity, and seasonality into tourism recommender systems. Information Technology & Tourism 27, 1 (2025), 189–226. [5] Ashmi Banerjee, Tunar Mahmudov, and Wolfgang Wörndl. 2024. Green Desti- nation Recommender: A Web Application to Encourage Responsible City Trip Recommendations. In Adjunct Proceedings of the 32nd ACM Conference on User Modeling, Adaptation and Personalization. 486–490. [6]Ashmi Banerjee, Adithi Satish, Fitri Nur Aisyah, Wolfgang Wörndl, and Yashar Deldjoo. 2025. Collab-REC: An LLM-based Agentic Framework for Balancing Recommendations in Tourism. arXiv preprint arXiv:2508.15030 (2025). [7]Ashmi Banerjee, Adithi Satish, Fitri Nur Aisyah, Wolfgang Wörndl, and Yashar Deldjoo. 2025. SynthTRIPs: A Knowledge-Grounded Framework for Benchmark Data Generation for Personalized Tourism Recommenders. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 3743–3752. [8] Oren Barkan, Veronika Bogina, Liya Gurevitch, Yuval Asher, and Noam Koenig- stein. 2024. A counterfactual framework for learning and evaluating explanations for recommender systems. In Proceedings of the ACM Web Conference 2024. 3723– 3733. [9]Keping Bi, Qingyao Ai, and W Bruce Croft. 2021. Asking clarifying questions based on negative feedback in conversational search. In Proceedings of the 2021 ACM SIGIR International Conference on Theory of Information Retrieval. 157–166. [10]Chainlit. 2026. Chainlit: Get Started – Overview. https://docs.chainlit.io/get- started/overview. [11] Philip Drammeh. 2025. Multi-Agent LLM Orchestration Achieves Determin- istic, High-Quality Decision Support for Incident Response. arXiv preprint arXiv:2511.15755 (2025). [12]Jiabao Fang, Shen Gao, Pengjie Ren, Xiuying Chen, Suzan Verberne, and Zhaochun Ren. 2024. A multi-agent conversational recommender system. arXiv preprint arXiv:2402.01135. [13] FastAPI. 2026. FastAPI Documentation. https://fastapi.tiangolo.com/. [14]Firebase and Google Cloud. 2026. Cloud Firestore Documentation. https://firebase. google.com/docs/firestore. [15] Google Cloud. 2026. Cloud Run Documentation. https://cloud.google.com/run. [16]Google Cloud. 2026. Overview of Agent Development Kit. https://docs.cloud. google.com/agent-builder/agent-development-kit/overview. [17] Google Cloud. 2026. Vertex AI Platform. https://cloud.google.com/vertex-ai. [18] Google DeepMind. 2024. Gemini 2.5: Technical Report. Technical Report. Google DeepMind. https://storage.googleapis.com/deepmind-media/gemini/gemini_ v2_5_report.pdf [19]Shengyu Gu. 2024. A survey of large language models in tourism (Tourism LLMs). Preprint on Qeios (2024). [20]Riccardo Guidotti. 2024. Counterfactual explanations and how to find them: literature review and benchmarking. Data Mining and Knowledge Discovery 38, 5 (2024), 2770–2824. [21]Haya Halimeh and Oliver Müller. 2025. Towards Greener Choices: Decision Information Nudging for Sustainability-Aware Recommender Explanations. In International Workshop on Recommender Systems for Sustainability and Social Good. Springer, 27–42. [22]Chengkai Huang, Junda Wu, Yu Xia, Zixu Yu, Ruhan Wang, Tong Yu, Ruiyi Zhang, Ryan A Rossi, Branislav Kveton, Dongruo Zhou, et al.2025. Towards agentic recommender systems in the era of multimodal large language models. arXiv preprint arXiv:2503.16734 (2025). [23]Ankur Joshi, Saket Kale, Satish Chandel, and D Kumar Pal. 2015. Likert scale: Explored and explained. British journal of applied science & technology 7, 4 (2015), 396–403. [24]Sara Kemper, Justin Cui, Kai Dicarlantonio, Kathy Lin, Danjie Tang, Anton Ko- rikov, and Scott Sanner. 2024. Retrieval-augmented conversational recommen- dation with prompt-based semi-structured natural language state tracking. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2786–2790. [25]Reza Yousefi Maragheh and Yashar Deldjoo. 2025. The Future is Agentic: Defini- tions, Perspectives, and Open Challenges of Multi-Agent Recommender Systems. arXiv preprint arXiv:2507.02097 (2025). [26]Noemi Mauro, Livio Scarpinati, Fabio Ferrero, Angelo Geninatti Cossatin, and Claudio Mattutino. 2024. Point-of-Interest Recommender Systems: Nudging towards Sustainable Tourism. In Adjunct Proceedings of the 32nd ACM Conference on User Modeling, Adaptation and Personalization. 491–495. [27]Qiyao Peng, Hongtao Liu, Hua Huang, Qing Yang, and Minglai Shao. 2025. A survey on llm-powered agents for recommender systems. arXiv preprint arXiv:2502.10050 (2025). [28]Xuhui Ren, Hongzhi Yin, Tong Chen, Hao Wang, Zi Huang, and Kai Zheng. 2021. Learning to ask appropriate questions in conversational recommendation. In Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval. 808–817. [29]Ivan Sekulić, Mohammad Aliannejadi, and Fabio Crestani. 2021. Towards Facet- Driven Generation of Clarifying Questions for Conversational Search (ICTIR ’21). Association for Computing Machinery, New York, NY, USA, 167–175. doi:10. 1145/3471158.3472257 [30]Ivan Sekulić, Weronika Łajewska, Krisztian Balog, and Fabio Crestani. 2024. Estimating the usefulness of clarifying questions and answers for conversational search. In European Conference on Information Retrieval. Springer, 384–392. [31]Zijian Shao, Jiancan Wu, Weijian Chen, and Xiang Wang. 2025. Personal Travel Solver: A Preference-Driven LLM-Solver System for Travel Planning. In Proceed- ings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 27622–27642. [32]Juntao Tan, Shuyuan Xu, Yingqiang Ge, Yunqi Li, Xu Chen, and Yongfeng Zhang. 2021. Counterfactual explainable recommendation. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management. 1784–1793. SIGIR ’26, July 20–24, 2026, Melbourne, VIC, AustraliaAshmi Banerjee, Adithi Satish, Wolfgang Wörndl, and Yashar Deldjoo [33]Ke Wang, Shuai Yan, Haoran Yuan, Yanling Huang, Yuhang Wu, Fei Li, Shengying Yang, and Huan Deng. 2025. Toward Interpretable and Persistent Personalization: A Memory-Augmented Agent Framework for LLM-Based Travel Planning. IEEE Access 13 (2025), 193125–193141. [34]Xiangmeng Wang, Qian Li, Dianer Yu, Qing Li, and Guandong Xu. 2024. Coun- terfactual explanation for fairness in recommendation. ACM Transactions on Information Systems 42, 4 (2024), 1–30. [35]Zhefan Wang, Yuanqing Yu, Wendi Zheng, Weizhi Ma, and Min Zhang. 2024. Macrec: A multi-agent collaboration framework for recommendation. (2024), 2760–2764. [36]Dianer Yu, Qian Li, Xiangmeng Wang, Qing Li, and Guandong Xu. 2023. Coun- terfactual explainable conversational recommendation. IEEE Transactions on Knowledge and Data Engineering 36, 6 (2023), 2388–2400. [37]Hamed Zamani, Susan Dumais, Nick Craswell, Paul Bennett, and Gord Lueck. 2020. Generating clarifying questions for information retrieval. In Proceedings of the web conference 2020. 418–428.