Settings

Everything is stored in this browser. The defaults are tuned with the evaluation set, so most people only need the first two sections.

Language model

Generates answers from retrieved passages. Retrieval, citations and evaluation work without it.

Connection

How the browser reaches a model.

Automatic (recommended)

Uses the model configured on the server when one is reachable, and otherwise the in-browser model — so answers are always generated, even on a deployment with no model of its own.

In-browser model (free and private)

A small open-weight model runs in this tab on your GPU (or CPU). Downloaded once, then cached by the browser; prompts and documents never leave your device.

Server provider

The Next.js API route calls the provider set in environment variables (Ollama locally, or an optional hosted API). Keys never reach the browser.

Ollama on this computer (private mode)

The browser talks directly to Ollama on your machine — the fastest option if you run it, and prompts stay on your computer.

In-browser model

Runs on your device. Larger models answer better but take longer to download and run.

Recommended: cited every answer in our benchmark, invented no numbers and refused what the documents do not cover.

1.2B parameters · LFM Open License v1.0

No WebGPU — will run on the CPU (slower) · 850 MB download on first use

Status

Checking…

Temperature · 0.2

Lower = more literal and repeatable. Grounded answers work best at 0.1–0.3.

Strict grounding

Refuse to answer (without calling the model) when retrieval finds no relevant evidence. Recommended.

Privacy & data

What stays on your device and what does not.

Stays in this browser

Your uploaded files (parsed in the browser; the original file is not kept)
Extracted text, chunks and embedding vectors (IndexedDB)
Chats, mock interviews, question bank and analyses

Leaves the browser

For each question: the prompt with the top passages is sent to the language model. With local Ollama, it never leaves your computer; with a hosted provider it goes to that provider.
Model downloads: weights from Hugging Face and runtime files from jsDelivr (no document data).

There are no accounts, analytics or server-side storage of documents. Clearing your browser data removes everything.