FAQ¶

Q. What are the system requirements and recommended device specifications?¶

A. The required device configuration depends on the expected inference latency and usage pattern (CPU-based inference, GPU-based inference, etc.). Below are the system requirements recommended by our company.

Minimum Requirements¶

Software Requirements¶

Item	Requirement
Supported OS	Windows / macOS / Linux

Hardware Requirements¶

Item	Requirement
Memory	8 GB or more
CPU Architecture	Intel / ARM
CPU Performance	1 GHz, 4 cores or more
Storage	At least 2 GB of free disk space

GPU Requirements¶

Item	Requirement
GPU Requirement	Not required (can run on CPU only)
GPU Architecture	NVIDIA (if GPU is used)
GPU VRAM	8 GB or more (if GPU is used)

Network Requirements¶

For offline device-authenticated licenses

Item	Requirement
Runtime	Not required (can operate in an offline environment)
Initial Setup	Network connection required

Recommended Requirements (YomiToku)¶

Software Requirements¶

Item	Requirement
Supported OS	Windows / macOS / Linux

Hardware Requirements¶

Item	Requirement
Memory	16 GB or more
CPU Architecture	Intel / ARM
CPU Performance	2 GHz, 16 cores or more
Storage	At least 2 GB of free disk space

GPU Requirements¶

Item	Requirement
GPU Requirement	Recommended
GPU Architecture	NVIDIA
GPU VRAM	16 GB or more

Network Requirements¶

For offline device-authenticated licenses

Item	Requirement
Runtime	Not required (can operate in an offline environment)
Initial Setup	Network connection required

Q. What information is transmitted during online authentication and execution logging?¶

A. Only the following information is transmitted during communication.

Information Sent During License Authentication¶

{
    "check_timing": "launch",
    "license_key": "string",
    "machine_fingerprint_hash": "string",
    "secret_key": "string"
}

machine_fingerprint_hash is sent only when using a perpetual (one-time purchase) license.
Actual device information is not transmitted; only a hashed value is used.

Information Sent in Execution Logs¶

{
    "attributes": {
        "modules": []
    },
    "count": 0,
    "license_key": "string",
    "timestamp": "2019-08-24T14:15:22Z"
}

The contents of processed images or PDF files are not transmitted.
Only information required for usage aggregation and license management is sent.

Q. Japanese fonts are not displayed correctly or are not rendered in images when processing PDFs.¶

A. Japanese fonts may not be installed in your execution environment. Please install commercially usable Japanese fonts such as IPA fonts.

Linux¶

apt-get update && \
apt-get install -y \
  fonts-ipafont \
  fonts-ipaexfont \
  fonts-noto-cjk

macOS¶

brew tap homebrew/cask-fonts
brew install --cask font-ipaexfont
brew install --cask font-noto-sans-cjk
brew install --cask font-noto-serif-cjk

Windows¶

Visit https://moji.or.jp/ipafont/
Extract the ZIP file
Right-click the .ttf files and select Install

Q. How can I check the number of processed pages?¶

A. You can check usage via the API or the dashboard.

Using the API¶

query_count \
  --license_key ${YOMITOKU_LICENSE_KEY} \
  --secret_key ${YOMITOKU_SECRET_KEY}

Using the Dashboard¶

Available for subscription licenses (Standard plan or higher).
Dashboard accounts are issued separately and will be provided via email.

Q. How do I use the orientation correction module? What should I do if images are rotated sideways?¶

A. Automatic correction is available using the orientation correction AI.

CLI¶

Specify the --rotate_detection option.

yomitoku ${path_data} --rotate_detection

Python API¶

Set enable_preprocess=True.

analyzer = DocumentAnalyzer(enable_preprocess=True)

Q. I would like to use YomiToku as a Web API.¶

A. YomiToku-Pro includes a built-in REST API server. Deployment via Docker is also supported.

See the Server page for details.

Q. During installation, the following message appears and the package name is shown as `UNKNOWN`.¶

Successfully built UNKNOWN
Installing collected packages: UNKNOWN
Successfully installed UNKNOWN-0.0.0

A. The version of build tools such as setuptools may be outdated. Please update them using the following command.

python -m pip install --upgrade pip setuptools>=77 wheel build

Q. CUDA is not recognized on newer GPUs such as the RTX 50 series. How do I install a CUDA-enabled build of PyTorch?¶

A. Newer GPUs such as the RTX 50 series may require a newer PyTorch build (e.g., one built for CUDA 12.8) than the version installed by default through the package dependencies. Install the dependencies first, then replace only the PyTorch-related packages with the official CUDA builds.

Installation Steps (using uv)¶

After cloning the repository, install the dependencies.

uv sync --extra gpu

Then install the PyTorch packages that match your CUDA version (example: CUDA 12.8).

Windows (PowerShell)Linux

uv pip install --upgrade `
  torch==2.7.0 `
  torchvision==0.22.0 `
  --index-url https://download.pytorch.org/whl/cu128

uv pip install --upgrade \
  torch==2.7.0 \
  torchvision==0.22.0 \
  --index-url https://download.pytorch.org/whl/cu128

Warning

Re-running uv sync afterwards will revert PyTorch to the version declared in pyproject.toml. In that case, perform the replacement step again.

Installation Steps (using pip)¶

After installing YomiToku-Pro, replace only the PyTorch-related packages in the same way.

pip install --upgrade torch==2.7.0 torchvision==0.22.0 --index-url https://download.pytorch.org/whl/cu128

Verifying CUDA Availability¶

After installation, verify that CUDA is recognized with the following command.

uv run python -c "import torch; print(torch.__version__, torch.version.cuda, torch.cuda.is_available())"

You can confirm that inference is running on the GPU using the following methods:

Monitor GPU utilization and memory usage with nvidia-smi during execution
Check CUDA memory allocation with torch.cuda.memory_allocated() and similar APIs

Note that if device="cuda" is specified while CUDA is unavailable, YomiToku-Pro does not silently fall back to the CPU; an error is raised at initialization or inference time. PDF rasterization, pre/post-processing, and Markdown generation always run on the CPU, so CPU usage during execution does not by itself indicate a CPU fallback of the inference models.

Notes¶

Use the latest official NVIDIA driver compatible with your CUDA version and GPU. YomiToku-Pro does not pin a specific driver version.
YomiToku-Pro uses only the standard public APIs of PyTorch / torchvision, and this kind of minor version update is generally compatible. However, since this configuration differs from the versions declared in the package dependencies, we recommend validating with your actual documents and PDFs before applying it to production workloads.

FAQ¶

Q. What are the system requirements and recommended device specifications?¶

Minimum Requirements¶

Software Requirements¶

Hardware Requirements¶

GPU Requirements¶

Network Requirements¶

Recommended Requirements (YomiToku)¶

Software Requirements¶

Hardware Requirements¶

GPU Requirements¶

Network Requirements¶

Q. What information is transmitted during online authentication and execution logging?¶

Information Sent During License Authentication¶

Information Sent in Execution Logs¶

Q. Japanese fonts are not displayed correctly or are not rendered in images when processing PDFs.¶

Linux¶

macOS¶

Windows¶

Q. How can I check the number of processed pages?¶

Using the API¶

Using the Dashboard¶

Q. How do I use the orientation correction module? What should I do if images are rotated sideways?¶

CLI¶

Python API¶

Q. I would like to use YomiToku as a Web API.¶

Q. During installation, the following message appears and the package name is shown as UNKNOWN.¶

Q. CUDA is not recognized on newer GPUs such as the RTX 50 series. How do I install a CUDA-enabled build of PyTorch?¶

Installation Steps (using uv)¶

Installation Steps (using pip)¶

Verifying CUDA Availability¶

Notes¶

Q. During installation, the following message appears and the package name is shown as `UNKNOWN`.¶