Retrieval-augmented generation (RAG) pipelines have become the de-facto approach for building AI assistants with external knowledge. Given a user query, RAG pipelines retrieve (R) information from external sources, before invoking a Large Language Model (LLM), augmented (A) with this information, to generate (G) responses. However, developing effective RAG pipelines is challenging because retrieval and generation components---often chained in varying orders---are intertwined, making it hard to identify which component(s) cause errors in the output. Developers often need to answer "what-if" questions---e.g., what if chunk sizes were larger or retrieval used embeddings versus keywords---but such experimentation requires hours of re-processing. We present RAGGY, a developer tool that enables rapid "what-if" analysis by combining a Python library of composable RAG primitives with an interactive debugging interface. We contribute the design and implementation of RAGGY, insights into expert debugging patterns through a qualitative study with 12 engineers, and design implications for RAG tools.
Conversational search, powered by Large Language Models (LLMs), has rapidly become a dominant mode of information seeking. While LLMs reduce the effort of information seeking, they also introduce the risk of distorted information deceptively embedded in responses. Prior work has sought technical mitigations, but such distortion cannot be fully eliminated. We therefore shift the focus to the user level, supporting users in deliberately engaging with information when reading LLM responses. We conducted a user study with frequent conversational search users (N=16), comparing a baseline with two probes—LLM-box (LLM-as-a-judge feedback) and Thinking-box (checkpoints from hallucination patterns)—to examine how these probes influenced users’ recognition of distorted information and their experience of guidance. Our findings indicate that even indirect suggestions significantly improved users’ ability to filter distorted information, while also revealing that guidance must be selective to prevent cognitive overload. These insights point to design implications that enable more deliberate user engagement with LLM responses.
When AI systems allow human-like communication, they elicit increasingly complex relational responses. Knowledge workers face a particular challenge: They approach these systems as tools while interacting with them in ways that resemble human social interaction. To understand the relational contexts that arise when humans engage with anthropomorphic conversational agents, we need to expand existing human-computer interaction frameworks. Through three workshops with qualitative researchers, we found that the fundamental ontological and relational ambiguities inherent in anthropomorphic conversational agents make it difficult for individuals to maintain consistent relational stances toward them. Our findings indicate that people's articulated positioning toward such agents often differs from the relational dynamics that occur during interactions. We propose the concept of relational dissonance to help researchers, designers, and policymakers recognize the resulting tensions in the development, deployment, and governance of anthropomorphic conversational agents and address the need for relational transparency.
Learning clothes repair is challenging for novices, who face interconnected procedural and embodied challenges, especially when learning alone. Existing tools fail to provide holistic support as interactive tutors and lack awareness of the embodied interactions of working with textiles. This paper presents a multi-phase study that investigates these challenges and explores the design space for a Human-Touch-Aware conversational agent (RepairBot). We began with an in-depth autoethnography to understand the novice experience, which informed the development of the RepairBot Conversation Framework (RBCF) together with a design implementation of a technology probe. Using the RepairBot prototype together with a Wizard-of-Oz approach to simulate Human-Touch-Awareness, we investigated how a conversational agent could support repair learning in novices as well as engage them with their own clothes-repairing projects. Subsequent lab and in-home studies with novice participants suggested specific conversational and embodied mechanisms that would facilitate novices' holistic understanding of repair, increase their confidence, and elicit attentive touch and emotional reflection. We bring these mechanisms together in the framework presented in this paper.
Caregivers often experience emotional difficulties and social isolation due to their demanding caregiving duties. Conversational AI has the potential to provide emotional support, yet it lacks effective emotional-regulation support. In this study, we conducted focus groups and semi-structured interviews with mental health professionals and caregivers (n = 17) to explore the potential benefits, challenges, and concerns of users on the applications of conversational AI for caregivers’ emotional support. Our findings suggest that, while current text-based conversational AI is deemed valuable for emotional support, there is a desire to have a more empathic AI, an AI that actively listens, takes cultural, religious, and linguistic context into consideration; and makes humans feel heard. We examined the dimensions of empathic AI in mental health, from authenticity and trust to over-reliance, misuse, and even exacerbating mental health problems, and how this can potentially be addressed to improve caregivers’ well-being.
Large language models (LLMs) enable conversational agents (CAs) to express distinctive personalities, raising new questions about how such designs shape user perceptions. This study investigates how personality expression levels and user-agent personality alignment influence perceptions in goal-oriented tasks. In a between-subjects experiment (N=150), participants completed travel planning with CAs exhibiting low, medium, or high expression across the Big Five traits, controlled via our novel Trait Modulation Keys framework. Results revealed an inverted-U relationship: medium expression produced the most positive evaluations across Intelligence, Enjoyment, Anthropomorphism, Intention to Adopt, Trust, and Likeability, significantly outperforming both extremes. Personality alignment further enhanced outcomes, with Extraversion and Emotional Stability emerging as the most influential traits. Cluster analysis identified three distinct compatibility profiles, with "Well-Aligned" users reporting substantially positive perceptions. These findings demonstrate that personality expression and strategic trait alignment constitute optimal design targets for CA personality, offering design implications as LLM-based CAs become increasingly prevalent.
What should an intelligent in-vehicle assistant (IVA) look like, and how should it behave to truly enhance the in-car experience? We present a large-scale video-based online experiment (n = 1238) exploring how IVA design factors influence user perceptions. Participants evaluated two scenarios (adjusting temperature, adjusting seat position) across 32 conditions varying in autonomy (user-initiated, system-initiated, autonomous with explanation, autonom- ous without explanation), embodiment (abstract virtual agent, humanlike virtual agent, abstract robot, humanoid robot), and conversational style (formal, informal). Contrary to prevailing academic trends, our findings reveal a clear preference against robotic embodiments and high levels of autonomy, sometimes even when explainable. Instead, participants favored proactivity with lower system autonomy and less anthropomorphic designs. We discuss how these insights challenge current design assumptions and offer concrete guidelines for shaping IVAs that align with driver expectations and comfort. This work contributes an empirically grounded understanding of IVA appearance, behavior, and communication style to inform future human-centered automotive interaction design.