# 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 from `GPU`; `audio-analyzer` and `text-to-speech` can also be pinned to `GPU` for 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: ```bash 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](https://dgpu-docs.intel.com). ## Host Packages The pull flow needs only Docker. Building from source and running `make benchmark` additionally need: ```bash sudo apt-get update sudo apt-get install -y git git-lfs python3-venv ffmpeg ``` | Package | Needed for | Symptom if missing | | --- | --- | --- | | `git`, `git-lfs` | Cloning this repo and the pinned upstream sources | Submodule or model assets fail to check out | | `python3-venv` | `make benchmark` builds a virtualenv for the benchmark orchestrator | `ensurepip is not available` | | `ffmpeg` | 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: ```bash 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 if `kiosk-ui` is exposed beyond localhost.