DataLens is a multimodal AI-powered data storytelling platform built for the Gemini Live Agent Challenge (Creative Storyteller category). It goes beyond traditional charts, transforming raw datasets into engaging, multi-perspective stories using Gemini's advanced interleaved output features.
- Live App: Try DataLens Live
- Tip: You can use the provided
spotify_tracks.csvas a sample dataset to test out the platform!
- Tip: You can use the provided
- Live Demo Video:
DataLens.Demo.Video.mov
- Google Cloud Deployment Proof:
Timeline.1.mov
- Upload & Ingest: Users upload CSV/Excel files or connect their database.
- Profiling: Pandas automatically profiles the dataset to build a strong
data_profilecontext. - Smart Searching (RAG): Vertex AI Vector Search indexes the dataset, stories, and external documents.
- Interleaved Story Generation:
gemini-3-prosimultaneously streams 3 distinct story formats with inlineIMAGE_PROMPTtags. - Parallel Image Generation: As text streams,
IMAGE_PROMPTtags trigger parallel Nano Banana Pro (gemini-3-pro-image-preview) calls, weaving images directly into the story. - Live Voice Agent: A bidirectional WebSocket connects users to the Gemini Live API, enabling conversational data exploration with full RAG context and on-demand image generation.
- Persistence: Sessions (stories, generated artifacts, conversation history) are automatically saved to Google Cloud Storage (GCS) and Firestore.
- Triple-Perspective Storytelling:
- 🧒 ELI5: Simple metaphors with fun, cartoon-style images.
- 🏢 Architecture: Schema designs and ER diagram images for engineers.
- 📊 Analyst: Business KPIs, trends, UI charts, and infographic elements.
- Interleaved Output: Real-time generation of textual narrative gracefully intertwined with dynamic image creation.
- Gemini Live Voice Agent: Hands-free, real-time chat with an intelligent "Voice Orb" that can reason over your datasets and instantly illustrate visual concepts.
- RAG-Powered Conversations: Embeddings seamlessly integrate external user documents, dataset structural stats, and previous story elements for highly contextual insights.
- Storyteller:
gemini-3-pro(Interleaved Output) - Image Generation:
gemini-3-pro-image-preview(Powered by Nano Banana Pro) - Voice Agent: Gemini Live API via bidirectional WebSocket (
google-adk) - Embeddings:
text-embedding-004(Vertex AI Vector Search)
- Framework: FastAPI (Python 3.11, fully async)
- Data Processing: Pandas, openpyxl
- Database & Storage: Google Cloud Storage (GCS), Firestore
- Vector DB: Vertex AI Vector Search (768 dims, Cosine distance)
- Deployment: Cloud Run (Dockerized)
- Framework: React 18 + TypeScript + Vite
- Styling & UI: Tailwind CSS, Framer Motion (streaming animations)
- State Management: Zustand
- Real-time Comms: Server-Sent Events (SSE) for stories, WebSockets for Live Agent
- Visualization: Recharts
datalens/
├── backend/ # FastAPI application, AI endpoints, and WebSocket agents
│ └── services/ # Core logic: gemini_story, live_agent, rag_pipeline
├── frontend/ # React UI, hooks for SSE & WebSockets, Dashboard layout
├── infrastructure/ # Terraform configs & deployment scripts for GCP
├── CLAUDE.md # Detailed Architecture Reference
├── TROUBLESHOOTING.md # Common issues and debugging instructions
└── deploy.sh # Deployment script for Google Cloud Run
Ensure you have Python 3.11+, Node.js, and a Google Cloud Project with the required APIs enabled.
- Clone the repo
- Backend Setup:
cd backend pip install -r ../requirements.txt # Set up your .env file based on .env.example python -m uvicorn main:app --reload
- Frontend Setup:
cd frontend npm install npm run dev
(See TROUBLESHOOTING.md for more detailed setup caveats.)
To deploy your own instance to Google Cloud Run, execute the deployment script:
./deploy.shRequires gcloud CLI authenticated and properly configured.
Built with ❤️ for the Gemini Live Agent Challenge.