Inspiration
Connecting concepts from various fields can lead to advancements in all aspects of life. Academic researchers often get stuck in their own domain's language and paradigms. By finding structural analogies—patterns that hold true across different fields—they can strengthen their arguments, find new ways to communicate complex ideas, and potentially discover cross-disciplinary solutions.
What it does
Rephrase helps academic researchers strengthen arguments in their papers by finding structural analogies from other academic fields.
Context Extraction: It parses an uploaded research paper (PDF or Word) to understand the domain, thesis, and keywords. Structural Analysis: It strips a specific claim or "point" of its domain-specific language to find its underlying abstract structure. Interdisciplinary Matching: It searches across fields like Mathematics, Physics, Biology, and Economics to find similar structural patterns. Evidence Procurement: It integrates with the OpenAlex API to find real-world academic citations that support the identified analogy. Argument Reframing: It uses AI to rewrite the researcher's original point using the terminology of the analogous field, providing a fresh perspective and a "how to use this" guide for inclusion in their paper. Visual Exploration: It presents these connections in an interactive 3D Force Graph, where the link distance represents the "structural strength" of the analogy.
How we built it
Using React Frontend and FastAPI backend with various AI tools and endpoints that transform and summarise data using LLMs.
Frontend: Built with React and react-force-graph-3d for the interactive visualization. UI components are styled with custom CSS and Lucide icons. Backend: A Python FastAPI server handles the logic and orchestration. AI Engine: Powered by Gemini 3.0 (via Google's AI Studio/Vertex AI) for complex tasks like structural extraction, analogue generation, and evaluation. The system is designed with a provider pattern to also support OpenAI and Anthropic. Data Integration: Integrated with the OpenAlex API to retrieve real metadata and DOIs for academic papers. Package Management: Managed with uv for lightning-fast Python dependency handling and npm for the frontend.
Challenges we ran into
Structural Rigor: Ensuring that the AI identifies deep structural parallels rather than just thematic similarities was a major prompt engineering challenge. Session Management: Handling document state in memory across different API calls (upload vs. search) without requiring a complex database setup. 3D Visualization: Fine-tuning the 3D graph physics (like scaling link lengths by analogy strength) to make sure users can easily navigate between different field clusters. Data Consistency: Parsing diverse PDF formats to extract clean text for AI analysis while respecting word limits (6,000 words). Accomplishments that we're proud of Dynamic Reframing: The ability to translate a complex psychological point into the language of fluid dynamics or set theory is remarkably effective. Seamless Integration: Successfully connecting the abstract world of LLM generation with the concrete world of academic citations via OpenAlex. Interactive UI: A sleek, modern interface that makes "research surfing" feel intuitive and visually engaging.
What we learned
Prompt Engineering is Key: The difference between a "good" analogy and a "hallucination" depends entirely on how the structural pattern is defined and evaluated in the system prompt. Interdisciplinary Patterns: Building this tool reinforced how many core concepts in science and humanities share identical underlying mathematical or logical structures. Agentic Workflows: Even simple search pipelines benefit from "agentic" steps like extraction, evaluation, and refinement.
What's next for Rephrase
Agentic Search: Implementing a more robust agent that can iterate on search queries if the initial OpenAlex results are weak. Collaborative Graph: Allowing researchers to "save" analogies to a persistent graph that others can browse. Direct Citations: Integrating with Zotero or Mendeley for one-click citation exports. Semantic Search: Moving beyond keywords to full semantic search for the analogy candidates.
Log in or sign up for Devpost to join the conversation.