Skip to content

getting-started.md

Latest commit

 

History

History
99 lines (70 loc) · 3.13 KB

getting-started.md

File metadata and controls

99 lines (70 loc) · 3.13 KB

Getting started

At the end of this tutorial you'll have protoApp running locally with a streaming chat UI talking to an OpenAI-compatible server that's living inside the Tauri process.

No model download is required — the default build uses a streaming stub so you can verify the whole stack works before committing to a 2.5 GB model pull.

1. Prerequisites

You need:

  • macOS, Linux, or Windows
  • Rust 1.80+ (rustc --version)
  • Node.js 20+ and pnpm 9+
  • The platform prerequisites for Tauri 2

Verify:

rustc --version
pnpm --version
node --version

2. Clone and install

git clone https://github.com/protolabsai/protoApp
cd protoApp
pnpm install

3. Run the app

pnpm tauri dev

The first launch compiles the Rust workspace (~30 seconds clean). When the window opens, you'll see three tabs: Chat, Transcribe, and Speak.

4. Send a message

Type "hello" and press Send. You should see a streaming reply like:

[stub reply — build with --features llm (optionally with metal on macOS or cuda on NVIDIA, e.g. --features "llm metal") for real inference] You said: hello

The llm feature is the one that pulls in llama-cpp-2 and the default Qwen3-4B-Instruct-2507 model; metal and cuda are GPU backends that only matter once llm is on.

That "stub reply" is the point of this tutorial — it proves:

  1. The frontend OpenAI SDK resolved the Tauri command get_api_base_url and got a http://127.0.0.1:<port> URL.
  2. The Rust side bound an Axum server on an ephemeral port.
  3. /v1/chat/completions streamed Server-Sent Events back to the browser.
  4. The frontend accumulated deltas and rendered them live.

If any of those steps broke, it would have been obvious — no model to blame.

5. Kick the voice panels

Switch to the Transcribe tab, click Record, say something for a few seconds, click Stop. The stub STT echoes back the byte count of your clip so you can confirm the mic path, the multipart upload, and the server round-trip all work. Real Whisper transcription comes online with --features stt.

Switch to the Speak tab, type something, click Speak. You'll hear one second of silence — that's the valid WAV the server returns while the real Kokoro engine is still pending. The audio element lets you confirm playback works; real voice arrives with --features tts.

6. Run the tests

In a second terminal:

cargo test --workspace

You should see every test passing across protolabs-voice-core and protoapp-agent (counts will drift as the suite grows; what matters is "all green").

Note: cargo test --workspace compiles the vendored zeroclaw runtime (the agent). Run git submodule update --init first, or the protoapp-agent build will fail on the missing vendor/zeroclaw.

Next