CosmoGraph is an AI-powered research exploration tool that visualizes relationships between scientific publications using semantic embeddings and knowledge graphs.
It helps users discover, summarize, and navigate NASA’s bioscience research by turning complex text data into meaningful visual connections.

• Demo link - Click here

CosmoGraph enables a new way of exploring space biology research by combining:

• Semantic Search — find publications similar in meaning, not just keywords.
• AI Summarization — generate concise summaries for each relevant publication.
• Knowledge Graph Visualization — view interconnected concepts and relationships through interactive nodes and links.
• Filtering — refine results by tags, similarity, or relevance.

A user searches: “water on Mars”

CosmoGraph finds publications that discuss Martian hydration, planetary geology, or astrobiology experiments, even if “water” and “Mars” don’t appear together.

It displays:

• A summary panel with related papers and their abstracts
• A knowledge graph connecting topics like Water Cycle, Soil Analysis, Life Detection
• Clickable nodes showing related research clusters

• User enters a query (e.g. “photosynthesis in microgravity”).
• The backend uses Sentence-BERT to convert it into a numerical embedding (a list of 384 values representing meaning).

• The embedding is sent to a PostgreSQL function (match_publications) that finds the most semantically similar publications using cosine similarity.

CREATE OR REPLACE FUNCTION match_publications(
  query_embedding vector(384),
  match_threshold float,
  match_count int
)
RETURNS TABLE (
  id bigint,
  title text,
  abstract text,
  tags text[],
  url text,
  similarity float
)
LANGUAGE plpgsql
AS $$
BEGIN
  RETURN QUERY
  SELECT
    p.id,
    p.title,
    p.abstract,
    p.tags,
    p.pdf_url,
    1 - (p.embedding <=> query_embedding) AS similarity
  FROM publications p
  WHERE 1 - (p.embedding <=> query_embedding) > match_threshold
  ORDER BY p.embedding <=> query_embedding
  LIMIT match_count;
END;
$$;

Once results are retrieved, the backend uses a text summarization model (e.g., OpenAI or Gemini) to condense each publication’s abstract.

Tags from publications are used as nodes.

Publications that share tags form edges between those nodes.

Each node links to the related publications in the sidebar.

The graph and results are displayed using react-force-graph-2d.

Inside backend/.env:

SUPABASE_URL=your_supabase_url
SUPABASE_KEY=your_service_role_key
OPENAI_API_KEY=your_openai_key  # optional for summarization

cd backend
python -m venv venv
venv\Scripts\activate   # on Windows
# or
source venv/bin/activate # on Mac/Linux

pip install -r requirements.txt

python -m uvicorn main:app --reload

Visit 👉 http://127.0.0.1:8000/docs to see automatically generated FastAPI API documentation.

Install & Run

npm install
npm run dev

Open the app at 👉 http://localhost:5173

• Citation or author-based connections
• Multi-query comparison mode
• Graph-based summarization
• Downloadable visualizations
• User accounts with Supabase Auth

• Abhay Murali M
• Ajay Krishna D
• Niranjana Ajay
• Rohan Raj
• Rose Francis
• Thaariq Hassan R