What it does How it works Get started Screenshots
You have handbooks, policies, contracts or notes spread across dozens of files. Finding one fact means opening them one by one.
RAG reads them for you. Type a question the way you would ask a colleague, and it:
- finds the paragraphs in your files that are closest in meaning to the question,
- answers using only those paragraphs, in the language you asked in,
- shows its sources. Every claim carries a number like 1. Click it and you see the file, the page and the exact passage.
If the answer is not in your documents, it says so instead of guessing.
Why "RAG"? It stands for Retrieval-Augmented Generation: the AI first retrieves relevant passages from your files, then generates an answer from them, not from memory.
| Numbered citations link each sentence to the passage behind it. No more taking the AI's word for it. | With Ollama, everything runs on your own computer. Your files and questions never leave it. | Mix English and Turkish documents freely. Ask in either; the answer comes back in the same language. |
| Keep topics apart: one collection for HR, one for engineering notes. Each has its own models and settings. | Open the retrieval panel under any answer to see every passage that was found, how close it was and whether it was used. | Drag and drop uploads, light and dark themes, works on a phone, English or Turkish interface. |
When you add a file, it is split into short passages. Each passage is turned into a meaning fingerprint (an embedding): a list of numbers that captures what the text is about. Two sentences that say the same thing in different words get similar fingerprints.
When you ask a question, it gets a fingerprint too, and the closest passages are picked out. That is why asking "how much time off do I get?" finds a paragraph that talks about "20 days of paid annual leave", even though the words differ. The AI model then writes the answer from those passages alone.
Setup takes a few developer tools and is done once. After that, everything happens in the browser.
What you need
| Docker Desktop | Runs the database and file storage |
| uv | Installs Python and the app's dependencies |
| Node.js 22.12 or newer | Builds the interface, once |
| Ollama or an OpenAI API key | The AI models. Ollama is free and local, for example ollama pull bge-m3 and ollama pull qwen3.5:2b |
Install and run
-
Create the settings file and a secret key:
cp .env.example .env uv run python -c "from cryptography.fernet import Fernet; print(Fernet.generate_key().decode())"Paste the printed value after
APP_SECRET_KEY=in.env. -
Start the database and file storage:
docker compose up -d
-
Install dependencies and prepare the database:
uv sync uv run python -m app.cli migrate
-
Build the interface:
cd frontend && npm install && npm run build && cd ..
-
Start the app:
uv run uvicorn app.main:app
-
Open http://localhost:8000.
First steps in the app
- Go to Providers, add an Ollama or OpenAI connection and press Test connection.
- Add two models to it: an embedding model (finds passages) and an LLM (writes answers). Fetch models lists what is available.
- Go to Collections, create one and pick those two models.
- Drop your files on the Documents tab. When they show Ready, switch to Ask.
Want something to try it with? eval/docs/ has sample files: an employee handbook, a returns policy, Kubernetes and PostgreSQL notes.
- Built for one person on one computer. There is no login, and the app only listens on your own machine.
- With Ollama, documents and questions stay on your computer. With OpenAI, the passages found for a question are sent to OpenAI to write the answer.
- API keys are stored encrypted and only ever shown masked.
Ask a question. The answer cites its source; clicking a number highlights the passage.
Look behind the answer. Every passage the search found, how similar it was (the tick marks the minimum score), and whether it went into the answer.
Dark theme |
On a phone |
- Code rules and commands: AGENTS.md
- Measuring search quality: eval/README.md
- API reference while the app runs: http://localhost:8000/docs
- Tests:
uv run pytest, lint:uv run ruff check . - License: MIT







