SUMIRE-n

LLM-powered spreadsheet for managing everything.

Excel democratized calculation with =SUM(). SUMIRE-n democratizes "what was the status of that?" with =LLM().

Write =LLM(B2, "summarize this", D2) in a cell, press ▶, and the LLM reads your local files and writes the result back into the spreadsheet. No LLM knowledge required — just type into cells.

SUMIRE-n: Structured Unified Map for Interlinked References and Execution - Notebook

日本語版 README はこちら

---

Features

• =LLM() in cells — Run a local LLM (Ollama) from any cell, just like =SUM()
• =xLLM() in cells — Run a cloud LLM (Gemini API) for higher quality
• File references — Put a file path in a cell (file:///path/to/spec.md), and the LLM reads it automatically
• Multiple notebooks — Manage projects as separate .sumiren files
• Desktop view — See all notebooks at a glance (React Flow)
• Fully local — Your data stays on your machine. No cloud required (unless you use xLLM)

---

Quick Start

Prerequisites

Software	Version	Required
Node.js	18+	✅
Python	3.12+ (tested on 3.12)	✅
Ollama	0.18+	⚡ For local LLM (=LLM())
Gemini API Key	—	⚡ For cloud LLM (=xLLM())

You need either Ollama or a Gemini API key. Both are optional if you just want to explore the UI.

1. Clone & install

git clone https://github.com/kikyujin/open-sumiren.git
cd open-sumiren

# Frontend
cd frontend
npm install
cd ..

# Backend
cd backend
python3 -m venv .venv
source .venv/bin/activate    # Windows: .venv\Scripts\activate
pip install -r requirements.txt
cd ..

2. Set up LLM (choose one or both)

Option A: Local LLM with Ollama (recommended)

Install Ollama from ollama.com/download, then:

# Recommended models
ollama pull gemma3:27b     # 27B — needs 16GB+ RAM, best quality
ollama pull gemma3:12b     # 12B — needs 8GB+ RAM, good balance

Other Ollama models may work but are untested.

No configuration needed. Ollama runs on localhost:11434 by default.

Option B: Cloud LLM with Gemini API

Get a free API key from Google AI Studio, then:

cp backend/.env.example backend/.env
# Edit backend/.env and set your key:
# GEMINI_API_KEY=AIza-xxxxxxxx

To use Gemini for all LLM calls (no Ollama needed):

# In backend/.env
GEMINI_API_KEY=AIza-xxxxxxxx
LLM_BACKEND=gemini

3. Run

./run.sh

Or start each service separately:

# Terminal 1: Backend
cd backend
source .venv/bin/activate
uvicorn main:app --host 0.0.0.0 --port 9300 --reload

# Terminal 2: Frontend
cd frontend
npm run dev

Open http://localhost:5173 in your browser.

4. First steps

Note: The UI is currently in Japanese. An English locale is planned for a future release.

When you open SUMIRE-n for the first time, you'll see an empty desktop. Here's how to get started:

1. Right-click the background → "📓 New notebook" → name it (e.g. my-project)
2. Double-click the notebook card to open the spreadsheet
3. Try this in the cells:

	A	B	C
1	hello world	translate in Chinese
2		=LLM(A1, B1, C1)

4. Select cell B2 (the formula cell) and click ▶ LLM実行 (top right) — C1 shows: 你好世界
5. Cmd+S (or Ctrl+S) to save

Next: Try putting a file path in a cell (e.g. file:///Users/you/project/README.md) and use =LLM() to summarize it.

---

How it works

Writing LLM formulas

=LLM(input, "prompt", output)

Part	Description
input	Cell reference(s) containing data or file paths. Multiple inputs OK
"prompt"	What to ask the LLM. Use a cell reference for Japanese text
output	Cell where the result goes. The formula stays so you can re-run

Examples:

=LLM(B2, "summarize in one line", D2)     — Summarize B2, write to D2
=LLM(B2, B5, C2, D2)                       — Multiple inputs (B2, B5), prompt in C2, output D2
=xLLM(B2, "translate to English", D2)      — Same syntax, uses Gemini API

File references: If a cell contains a file path like file:///Users/you/project/README.md, the file content is automatically loaded when the LLM runs.

Japanese prompt tip: Due to a Univer v0.18 bug, write Japanese prompts in a separate cell and reference it, instead of putting Japanese text directly in the formula.

Desktop & Notebooks

• Desktop (/): See all notebooks as cards. Right-click to create, rename, delete
• Sheet (/notebook/{name}): Spreadsheet view. Write formulas, press ▶ to run LLM
• Each notebook saves as a .sumiren file in the data/ directory

---

Environment Variables

Variable	Default	Description
GEMINI_API_KEY	—	Gemini API key. Required for =xLLM()
LLM_BACKEND	ollama	Backend for =LLM(). Set to gemini to use Gemini for everything

---

Using a different cloud LLM

SUMIRE-n uses Gemini by default for =xLLM(), but you can swap it for any LLM provider by editing a single file: backend/xllm_adapter.py.

The file has detailed comments explaining how to swap in OpenAI, Anthropic Claude, or any other provider. The only rule: keep the function signature generate_xllm(prompt, context) → str.

---

Tech Stack

Layer	Technology
Spreadsheet UI	Univer 0.18 (canvas-based, Apache-2.0)
Desktop view	React Flow v12
Frontend	React 19, Vite 8, TypeScript
Backend	FastAPI 0.115, Python 3.12
Local LLM	Ollama
Cloud LLM	Google Gemini API (REST)

---

Project Structure

open-sumiren/
├── frontend/           # React + Univer + Vite
│   └── src/
│       ├── pages/      # Desktop.tsx, Sheet.tsx
│       └── components/ # NotebookNode.tsx
├── backend/            # FastAPI
│   ├── main.py         # API endpoints
│   ├── llm_adapter.py  # Ollama adapter
│   └── xllm_adapter.py # Gemini adapter (swap this for other providers)
├── data/               # Notebook files (.sumiren)
└── run.sh              # Start everything

---

Roadmap

• Phase 0: LLM in cells (Univer + FastAPI + Ollama)
• Phase 1: Multiple notebooks, desktop view, cloud LLM, OSS release
• Phase 2: Chain execution, health checks, headline values
• Phase 3: Cross-notebook cell references, edge visualization

See docs/roadmap.md for details.

---

License

Apache License 2.0. See LICENSE.

---

Credits

Built by @kikyujin with 🦊 Elmar.