architecture diagram

Search4Cure Live

Inspiration

The biomedical research landscape is vast and constantly evolving. Researchers often struggle to find relevant information quickly, especially in areas like diabetes, where the literature grows daily. Our inspiration came from the need to accelerate discovery and help scientists navigate complex datasets, including papers, gene networks, and experimental results, in a meaningful way.

What it does

Search4Cure Live is an AI-powered biomedical research assistant. It enables researchers to:

Perform semantic and hybrid searches across biomedical literature.
Explore relationships between entities using graph-based queries.
Retrieve figures, tables, and specific sections from papers.
Quickly identify gene-disease associations and research methods.

All this is wrapped in an interactive web interface powered by ADK Web, with support for Graph RAG and hybrid search methods.

How we built it

Agent Layer: Built with Google ADK, using a custom Agent to manage biomedical search tools.
Search Tools: Implemented semantic, hybrid, graph, page, image, and table search capabilities.
Embeddings: Integrated Vertex AI embeddings (textembedding-gecko-001) for semantic understanding.
Graph Database: Used Neo4j-style graph queries to represent relationships between genes, diseases, and methods.
Web Interface: Exposed via adk web with FastAPI endpoints for session-based interactions.

Challenges we ran into

Embedding Model Errors: Initially faced 400 Model not found: TextEmbeddings. Resolved by switching to Vertex AI embedding models.
Graph Representation: Structuring biomedical entities in a meaningful graph required careful normalization of gene and disease names.
Real-Time Querying: Optimizing hybrid and semantic searches for speed without sacrificing accuracy was a key challenge.

Accomplishments that we're proud of

Successfully integrated Graph RAG and hybrid search for biomedical queries.
Built a robust web interface that allows researchers to explore papers, tables, figures, and graphs seamlessly.
Achieved accurate retrieval of complex biomedical relationships, supporting advanced research questions.

What we learned

Effective biomedical search requires domain-specific embeddings for high accuracy.
Graph-based approaches are powerful for exploring relationships, but require careful data curation.
Combining multiple search strategies (semantic, hybrid, graph) significantly improves retrieval relevance.

What's next for Search4Cure Live

Extend coverage to other diseases and conditions, beyond diabetes.
Incorporate multi-modal embeddings for images, tables, and figures.
Add personalized researcher profiles for saving queries and tracking discoveries.
Explore deployment on Vertex AI and cloud infrastructure for larger-scale access.