Constellation

The research tool that allows you to navigate to the edge of human knowledge by exploring your curiosities.

Inspiration

In a world of vast knowledge and rapid innovation, learning what research to conduct isn’t actually as intuitive as it seems. There is no apparent path for students, builders, and researchers to scope out avenues for cutting edge advancement. Even after constantly reading papers and skimming dozens of abstracts, we are unable to discern the ideas that truly matter. We lose foundational works, emerging breakthroughs, and logical advancements in topics we’re invested in.
That’s why we built Constellation to disrupt this.
Instead of forcing you to search blindly, it transforms a single topic or article into an interactive universe of related research. Every node explains why it matters, how it connects, and where to explore next, which helps us turn curiosity into real research initiative.

The Problem

Starting research is unnecessarily hard because there is no clear path from curiosity to a solid set of papers. Researchers spend hours skimming abstracts and jumping between tabs, but still don’t know what is foundational, what is truly relevant, or what to read next. That confusion stalls real research progress before it even starts.
Modern research tools don’t understand you. And because of that, they often aren’t useful.

Our Solution

We created Constellation, which is an AI-powered research exploration app.
A system that turns one topic into a navigable graph of knowledge, helping users move from curiosity to action.

What It Does

A Research Graph That Understands Your Exploration

Constellation begins with a topic or article URL and identifies a foundational paper. From there, users expand and traverse nodes to discover related work, forming a constellation. The constellation continues to grow until it reaches a “Frontier” node, which represents the most recent and relevant findings in that field.
Nodes are draggable, expandable, and organized by depth, allowing users to visualize the structure of a research field instead of scrolling through endless search results.

Deep Discovery Going Beyond Simple Search

Each node expansion uses a multi-stage AI/RAG pipeline:

Exa for semantic paper and web search
Google Gemini for query rewriting, paper selection, and frontier evaluation
Supermemory for retrieval-augmented generation (RAG) across stored papers

Interactive Research with RAG

We attempt to transform paper reading into a conversation.

Per-paper chat: Ask questions grounded in a specific paper’s content
Global search: Access a large knowledge base across all papers in a constellation with Supermemory and find connections between different fields.

Research becomes interactive, enabling deeper understanding without leaving the graph.

Immersive Visualization

Our UI models the research process as space exploration

Starfield landing page that is implemented with a graph view
Animated edges that show relationships between different papers
Minimap UI for large constellations
Frontier points that represent the cutting-edge of a topic

Exploration feels as novel as traversing a universe, helping to spark creativity and interest in any topic.

How We Built It

Constellation is a full-stack web application designed for real-time exploration and scalable knowledge retrieval.

Frontend

Next.js 16 (App Router), React 19
Tailwind CSS with shadcn-style components
Canvas API for starfield rendering and node visualization

Backend

Next.js server actions for search, RAG, and pipeline triggers
Python FastAPI pipeline for PDF ingestion and processing

Data & Persistence

Supabase (Postgres) for individual node data and storing graph state
Supermemory for chunked research paper content and semantic retrieval

External APIs

Exa for web search
Google Gemini for query rewriting, paper selection, PDF to Markdown conversion, and RAG integration with Supermemory for responses

Pipeline

FastAPI service: download PDF → convert to Markdown with Gemini → store in Supabase → index and embed in Supermemory

Challenges We Ran Into

Frontend: Making the graph feel full without lag

Rendering a starfield with an interactive node graph that included animations and a minimap had the potential to lead to a poor and slow user experience, especially as constellations grow to extremely large sizes. We had to carefully manage redraw frequency, make sure our states were lightweight, and avoid unnecessary rerenders so our final product was as smooth as possible.

Inference latency when expanding nodes

“Expand node” is a multi-step flow, as we have to rewrite the intent into a search query, search papers, select the best options, then potentially evaluate whether the paper is on the “frontier.” Early on, this felt slow because each step was in series. We later reworked the pipeline to reduce slowdowns by cutting down prompts for models, add parallel calls with Gemini to speed up queries, and use failsafes so exceptions failed gracefully even when a pipeline step broke.

PDF ingestion reliability for markdown within Supermemory

The papers we extracted from were messy, often having multi-column layouts, figures, math blocks, and footnotes. If the PDF to markdown step became noisy this would directly lead to our retrieval suffering and getting worse as answers degraded. We had to chunk intelligently with Gemini and handle partial failures so the RAG system remains useful when faced with some blank data.

Multi-Service Orchestration

Coordinating Exa, Gemini, Supermemory, Supabase, and the ingestion pipeline required robust error handling and fallback strategies. For example, we used Gemini to optimize prompts for Exa’s web search capabilities and Supermemory to integrate RAG with Gemini for responses grounded in research papers. This enabled us to have a robust RAG/inference pipeline to power our constellation’s interactivity with chatbots.

Accomplishments We’re Proud Of

Carving a Path Towards Research Excellence

We are proud that we made research feel simple and guided instead of overwhelming. Constellation helps users start from one topic or paper and quickly build a clear path of related work, with reading and chat in the same place. This turns curiosity into real understanding faster.

Orchestrating raw intelligence.

The hard part wasn’t making calls to Gemini, but coordinating Exa + Gemini + Supermemory + Supabase + FastAPI. The result was a product that boasted meaningful expansion and grounded inference due to a robust AI pipeline.

Immersive UI that changes how you research

We’re proud that the starfield → collapse → constellation graph flow turns a normally boring workflow into a spatial experience where you can see your progress, follow branches, and stay oriented. We are proud that we were able to create a unique visualization for such a potentially dense and cluttered topic.

What We Learned

Personalization and innovation in research isn’t about spamming more papers, it’s about finding better pathways.
We learned that the real difficulty is reliability and structure. Expansion, ingestion, and retrieval must work together so results feel consistent, grounded, and useful instead of random. We also learned that a spatial, visual interface reduces overwhelm and keeps users oriented. This only works when users stay in control and the AI acts like a guide rather than an autopilot.