#group #LGresearch #AI #web | 🗓️ | Jun 2023 - Jun 2024 | | ------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Team. | Jiwon, Gwangho, Seungho ,Youjee, Gunhee | | Credit. | -Jiwon: Project management, Service planning, QA<br>-Gwangho: Back-end develop, Prompt engineer, QA<br>-Seungho: Service planning<br>-**Youjee: Service planning, Design lead, UX, Prototype, GUI, QA**<br>-Sinae: GUI, Design system<br>-Gunhee: Front-end develop | | About. | Collaboration with LG AI research team, having their AI technology. we made web platform for professional by using technology. it's been developed 3times including last version3. |  ##### _An AI platform that predicts user questions and stores responses as reports._ This platform was created in collaborated with LG research and our team. It is a professional conversational AI platform designed to minimize AI hallucinations by using documents related to the questions to generate evidence-based responses. The project was conducted in two phases. In the first phase, the focus was on generating responses based on conversations with AI. and on the right, there are reference related with the prompt. and we got feedbacks After the users tried it themselves, we gathered feedback and found that they had to ask several questions to get the information they wanted, which was inconvenient. They wanted to obtain various information with fewer questions. So In the second phase, we designed the responses to be more **document-oriented**, allowing users to gain more information like Chart, News and Subagent answers without needing to ask additional questions.  I participated in the planning process and led the UX/UI design. I also created the design system, referencing LG’s existing branding. Since the platform is intended for use by LG researchers and experts, the design concept was kept simple, while also considering LG's branding. We tried to reflect the reliability and professionalism of AI in the design. #### Interface  ###### Side Menu > Since the answers asked by users are documented, we found a need for users to organize them into folders based on similar categories. Instead of just stacking the questions in a list, we added functionality that allows users to create and customize folders. > When main agent [[#Response to prompt(main agent)👇]] is added to a folder, the sub-page [[#Sub-agent Content]] is included as well. However, if only the sub-page is added, it remains solely within the folder. <br>  ###### Query input >Lair IDs stored in the database can be searched using "@" (e.g., @Tesla). <br>   ###### Result status > The user questions and status of answers (Success, Loading, Fail) are displayed. The reason for including these status values is that the time taken to generate answers varies slightly for each response (e.g., sub-agent responses, charts, references). A mechanism was needed to inform users whether the answer is still being generated or has been completed. If the user is not satisfied with the answer, they can generate a new one using the 'Retry' button. <br>  >You can modify the query on the result screen, and after that, a new report will be generated. <br>  ###### Sub-agent Content > Each content is organized into pages, and the included references are shown as tags, allowing users to see them in advance. Unlike the main agent page, the sub-agent is structured as a single-depth page. As its importance decreases, it is repositioned lower on the screen (below). <br> ###### Response to prompt(main agent) > Response to user prompts: We discovered a need for users to see the response first when they ask a question, so I chose to present the conclusion at the top. So the final answer to the user's question is placed at the top, with the supporting references positioned below.   > The references serve as the basis for the main answer, and charts and external links can be dragged and dropped into the query input area. This allows users to understand the content or request chart analysis without reading all the information. Additionally, all text on the main agent page can be edited, making it easy for users to integrate additional content and save or make it into a report. <br>  > When you click on a chart, you can view it in detail on a dialog, and references related to the chart are shown at the bottom. <br>  > An error occurs while loading content on the result screen. #### Dark mode     ##### Design system  --- We worked on versions 1.0 and 2.0 over a year including development, with extensive communication between the developers and the team. While there were some limitations in implementing certain features due to the ongoing technical development, I'm looking forward to the release of version 3.0.