Empirica

Interacting with 3D Graph
Launch page

Inspiration

Biomedical researchers spend 15% of an average week up to 50% or more during intense periods of their time on literature review. When working on biomedical research, we realized finding connections across 50+ papers was nearly impossible manually. We needed a way to see the forest, not just the trees. What if AI could not only read papers but actively discover new ones, build knowledge networks, and generate research hypotheses automatically? That's why we built Empirica.

What it does

Empirica transforms biomedical research with three core capabilities:

1. Agentic Research Workflows: Turn on "Agent Mode" and watch Empirica autonomously search PubMed, discover papers via Google Scholar, download PDFs, extract entities, and build knowledge graphs—all while you focus on analysis.

2. Interactive Knowledge Graphs: Visualize genes, diseases, drugs, and their relationships in stunning 2D/3D force-directed graphs. Click nodes to explore connections, filter by entity type, and navigate through your research visually.

3. AI-Powered Intelligence:

RAG-Enhanced Chat: Ask questions and get answers with exact paper + page citations
Discovery Lab: AI generates research hypotheses by analyzing graph structure and document content, showing confidence scores and supporting evidence

Every insight is traceable, persistent, and citation-backed.

How we built it

Backend (Python):

FastAPI for async REST API with OAuth 2.0 authentication
scispaCy for biomedical named entity recognition (80+ entity types)
PyMuPDF for PDF text extraction
NetworkX for graph construction and analysis
sentence-transformers for semantic embeddings
Custom RAG system with entity-aware chunking
SQLite for persistence (projects, chat history, hypotheses)

Frontend (TypeScript):

React 18 + Vite for blazing-fast development
react-force-graph for 2D/3D WebGL visualization
Three.js for advanced 3D rendering with particles and curved edges
Tailwind CSS + Lucide icons for modern UI
Axios for API communication

AI Integration:

Claude 3.5 Sonnet for natural language processing
Custom prompt engineering for hypothesis generation
RAG architecture combining semantic search with graph context

External APIs:

PubMed E-utilities for paper search
Google Scholar scraping for PDF discovery
PMC and DOI resolution for full-text access

Challenges we ran into

1. Graph Performance: Visualizing 500+ nodes with real-time physics simulation tanked performance. We optimized by implementing WebGL rendering, reducing particle counts, and adding smart chunking for large graphs.

2. RAG Context Windows: Initially, RAG retrieved too much context, hitting Claude's token limits. We implemented entity-aware chunking that prioritizes relevant sections and semantic ranking to surface the best 7 chunks.

3. Real-time Progress Updates: Background jobs ran asynchronously, making it hard to show live progress. We implemented a polling system with granular status updates and persistent job tracking.

4. OAuth Token Expiration: During long research sessions (10+ minutes), Google OAuth tokens would expire mid-process, crashing the workflow. We implemented optional authentication for status endpoints and graceful error handling.

Accomplishments that we're proud of

✅ End-to-end autonomous research pipeline - From query to knowledge graph in minutes, fully automated

Production-quality RAG system - Semantic search + entity-aware retrieval with precise citations (paper + page)

Stunning visualizations - 3D graphs with colored edges, animated particles, proper lighting, and smooth interactions rival commercial tools

Full persistence layer - Chat history and hypotheses survive across sessions, making research continuity seamless

Citation transparency - Every AI-generated answer links back to exact source documents and pages

Graph-aware hypothesis generation - Discovery Lab analyzes both document content AND network structure to find insights humans might miss

Real-time updates - Live progress tracking during agentic research (papers found, analyzed, entities extracted)

Clean architecture - Modular, typed codebase with clear separation between services, ready for scale

What we learned

Technical:

scispaCy's biomedical NER capabilities are incredible but require careful prompt engineering to maximize extraction quality
WebGL force-directed graphs need careful optimization—curved edges, particles, and lighting all impact frame rates
RAG isn't just "throw everything at the LLM"—entity-aware chunking and semantic ranking dramatically improve answer quality
Background job management in FastAPI requires thoughtful state management and polling strategies
Three.js lighting models make a huge difference in 3D visualization readability

Research Domain:

Biomedical research has VERY specific entity types (genes, proteins, diseases, drugs, pathways)
PubMed's API is powerful but Google Scholar often has better PDF availability
Relationship extraction is hard—co-occurrence is a decent baseline but pattern matching catches many more connections
Researchers care deeply about citations and provenance—"trust but verify" is critical

Product:

Visual feedback is everything—users want to SEE progress, not just wait
Persistence matters more than we thought—researchers return to projects over days/weeks
The "magic moment" is when Discovery Lab surfaces a hypothesis the user hadn't considered
Autonomous workflows need to show their work—users want to understand what the AI did

What's next for Empirica

Expand Research Domains:

Multi-domain NER models - Integrate specialized models for chemistry, physics, computer science, and social sciences beyond biomedicine
Domain-adaptive entity extraction - Let users choose research domain (clinical, genomics, drug discovery) for optimized NER performance
Custom entity training - Allow researchers to fine-tune models on their specific subdomain

Enhanced Collaboration:

Real-time multi-user editing - Multiple researchers annotating and discussing the same graph simultaneously via WebSockets
Export to academic formats - BibTeX, RIS, EndNote, and auto-generated literature review sections with proper citations
Author network analysis - Visualize collaboration patterns and identify key researchers in your field

Smarter AI:

Fine-tuned biomedical LLM - Train domain-specific model for better entity extraction and more accurate hypothesis generation
Literature review generator - Auto-compose structured academic review sections from your knowledge graphs
Experiment suggestions - Recommend methodologies and protocols based on successful papers in your graph

Scale & Performance:

PostgreSQL migration - Handle concurrent access and larger datasets efficiently
Graph versioning - Track how your knowledge evolves over time with diff visualization
Handle 10,000+ paper graphs - Streaming processing and smart pagination for massive research projects

Our vision: Make AI-accelerated research accessible to every scientist, turning months of literature review into hours of insight discovery.