System Requirements#
Hardware#
CPU: x86_64. Intel Core Ultra (Meteor Lake) or newer is recommended. Older Intel Core / Xeon processors will run the stack but may be slower on OpenVINO inference paths.
Memory: 32 GB RAM minimum. 64 GB recommended when running the LLM, reranker, ASR, and TTS together with a warm cache.
Disk: 60 GB free SSD space recommended for model assets, the Hugging Face cache, vector storage, generated audio, and per-session storage. NVMe is preferred for faster first-run model export.
GPU (optional): Intel integrated GPU (Meteor Lake or newer iGPU) or a supported discrete GPU exposed via
/dev/dri. The RAG LLM and reranker benefit most fromGPU;audio-analyzerandtext-to-speechcan also be pinned toGPUfor higher throughput.Microphone: Not required on the host — audio is captured by the browser via the Web Audio API and uploaded to
kiosk-core.
Operating System#
Ubuntu 22.04 LTS (validated) or a compatible Linux distribution with a recent kernel.
Docker Engine and Docker Compose v2 for container deployment.
For GPU acceleration on Linux: Intel/OpenVINO host GPU runtime installed on the host:
sudo apt-get install -y intel-opencl-icd intel-level-zero-gpu level-zero sudo usermod -aG render $USER
Verify with
ls /dev/dri/renderD*. If the packages are not found, follow the Intel GPU driver guide.
Host Packages#
The pull flow needs only Docker. Building from source and running
make benchmark additionally need:
sudo apt-get update
sudo apt-get install -y git git-lfs python3-venv ffmpeg
Package |
Needed for |
Symptom if missing |
|---|---|---|
|
Cloning this repo and the pinned upstream sources |
Submodule or model assets fail to check out |
|
|
|
|
Transcoding benchmark/replay audio to 16 kHz mono PCM |
Benchmark audio fixtures fail to load |
Results Directory Permissions#
kiosk-core runs in-container as uid 1000 and writes benchmark output
to the bind-mounted results/ directory. If your host user is not uid
1000, the container cannot write there, the benchmark reports zero
transactions even though every turn succeeded, and nothing is logged
about it. Check and fix with:
id -u # if this is not 1000
mkdir -p results && chmod 777 results
Network#
Outbound internet access on first run to download model assets from Hugging Face, unless models are pre-staged under the per-service
models/and.cache/directories.Inbound access on the host for the published TCP ports:
7860,8010,8011,8012,8020.
Browser#
Any modern Chromium-based browser or Firefox with permission to use the microphone for
http://127.0.0.1:7860.The browser must be able to reach
kiosk-core(http://127.0.0.1:8012) from the same machine, or a routable address ifkiosk-uiis exposed beyond localhost.