lsh2546/gemma-nextstep

NextStep

Turning Official Documents into Action with Gemma 4

NextStep is a local-ready prototype that transforms an official document into three evidence-linked actions, an ICS calendar event, a persistent checklist, and a privacy-safe help card.

Current prototype

This repository contains a fully working static vertical slice for the representative scenario:

• —Korean school notice upload
• —English or Vietnamese question/response mode
• —Three structured actions
• —Source evidence and confidence labels
• —.ics calendar file export
• —Locally saved checklist
• —Redacted multilingual help card

The static browser demo uses deterministic sample data so the product flow can be inspected without downloading model weights. The submission app in gemma_app.py sends the uploaded image to the instruction-tuned google/gemma-4-E4B-it model and validates its evidence-linked JSON output.

Run

Gemma-powered reference implementation

The free Static Space hosts the interactive UX preview and does not run Python or model weights. To reproduce the actual Gemma 4 inference path in a Kaggle Notebook or GPU Python environment, accept the model terms, authenticate to Hugging Face, and run:

pip install -r requirements.txt
python app.py

Public Static Space demo

The default Hugging Face Space configuration serves index.html. Locally, open it directly or serve the folder:

python -m http.server 8000

Then open http://localhost:8000.

Gemma 4 integration boundary

The implementation boundary is analyze_document() in gemma_app.py. It:

1. 1.Accepts a document image, user question, and response language.
2. 2.Calls Gemma 4 E4B multimodal inference.
3. 3.Requires quoted source evidence for each action.
4. 4.Downgrades unmatched action/evidence references to uncertain.
5. 5.Returns structured JSON matching the UI contract.

Hosted demo and local mode must be described separately. Do not claim fully private/offline operation until local inference has been verified on the target device.