DEMO
Interacting with LLMs can feel more tactile.
Here, if a user drags two statements closer together in a graph, the LLM is prompted to refine these ideas into a more cohesive argument. Unlike traditional linear text, where information is presented in a fixed order and often requires sequential reading, a modular format allows users to prioritize and rearrange content based on more personal relevance and context.
To better visualize relationships between sentences.
A similarity threshold toggle adds more personalization to visualizations. For example, a high similarity threshold reveals sentences that are more directly related, which helps us focus on precise connections and avoid clusters of loosely related ideas. A lower similarity threshold can help reveal broader thematic or conceptual connections across the text. This can be helpful for high-level brainstorming, synthesis, or uncovering patterns that wouldn’t otherwise be obvious.
PROCESS
Design and develop. (and repeat)
I first built out a few interactive prototypes in Figma to help me visualize what the tool would look like and how it'll work. My initial prototype included more ways to organize the text, including organize by tone and length, but given the time constraints, I decided to develop the basic function of organizing by relevance. On the development end, Jupiter is built using React for the frontend, D3.js for graph visualization, and Compromise (nlp) for sentence parsing. The backend uses Node.js with Express.js. It integrates OpenAI's API, using the 'text-embedding-ada-002' model for embeddings and 'gpt-3.5-turbo' for text revision. Other technologies include Axios for HTTP requests, CORS for cross-origin requests, and dotenv for environment variable management.
CONSIDERATIONS
Potential use cases.
I think that gestural interactions with LLMs could support more creative thinking in scenarios involving brainstorming or building layered arguments. A dynamic approach lets users rapidly explore alternatives and potentially uncover new perspectives. LLMs could serve a more collaborative role rather than largely responding to discrete prompts.
Writing tools can reflect diverse cognitive styles.
Modular structures enables information engagement to feel more intuitive and comprehensible. For example, a modular format allows people to organize information in ways that align with their cognitive styles, such as arranging by sequence, relevance, tone, or thematic clusters. Linear thinkers can structure nodes step-by-step, while associative thinkers may prefer to group ideas around themes. I want to explore what possibilities this flexibility opens up for users to experiment with various modes of organization, and ways for interfaces to more naturally adapt to these changes.
Jupiter
Historically, writing began as a practical solution for recording transactions and maintaining accounts (like early Mesopotamian cuneiform or Egyptian hieroglyphs). These inscriptions were not necessarily designed to convey nuanced ideas or dialogue, but to impose order, standardization, and permanence on information. However, as our needs for understanding and collaborating with complex ideas have evolved, so too must our methods of interacting with text. I wanted to explore ways for people to more visually perceive and revise ideas collaboratively with LLMs.
Jupiter allows you to input text and visually explore the relationships between sentences through an interactive graph. By adjusting the proximity of nodes, you can explore how sentences might be revised to better align contextually with nearby sentences.
TIMELINE
Oct 2024, 1 week
CONTRIBUTION
Design, Development
PROJECT TYPE
Solo Project