dev(narugo): fix the docs

    versions of these models for efficient inference, available at 

    `deepghs <https://huggingface.co/deepghs>`_ repositories.

    In addition to standard tagging, the models can identify anime character IP (Intellectual Property)

    associations. For example, if a character like "misaka_mikoto" is detected, the system can map

    this to the "toaru_kagaku_no_railgun" (A Certain Scientific Railgun) IP. All IP names follow

    Danbooru-style tag conventions for consistency with existing anime tagging systems.

    Example::

        >>> from imgutils.tagging.pixai import get_pixai_tags

        >>> # Get tags with default thresholds

        >>> # Get all tags in a single dictionary

        >>> all_tags = get_pixai_tags('path/to/image.jpg', fmt='tag')

        >>> print("All tags:", all_tags)

        >>> # Get IP information for detected characters

        >>> ips = get_pixai_tags('path/to/image.jpg', fmt='ips')

        >>> print("Detected IPs:", ips)

"""

import json

    """

    Load the tag metadata from Hugging Face Hub with caching.

    This function downloads and loads the CSV file containing tag names and categories

    for the specified model. The DataFrame contains columns for tag names, categories,

    and other metadata.

    This function downloads and loads the CSV file containing tag names, categories,

    and IP (Intellectual Property) associations for the specified model. The DataFrame

    contains columns for tag names, categories, and other metadata including character

    IP mappings when available.

    :param model_name: Name of the model

    :type model_name: str

    :return: DataFrame containing tag information with columns like 'name', 'category'

    :rtype: pd.DataFrame

    :return: Tuple containing (DataFrame with tag information, dictionary mapping character tags to their IPs)

    :rtype: Tuple[pd.DataFrame, Dict[str, List[str]]]

    """

    df_tags = pd.read_csv(hf_hub_download(

        repo_id=_get_repo_id(model_name),

          search after L2 normalization

        - ``logits``: raw 1-dimensional logits output from the model

        - ``prediction``: 1-dimensional prediction probabilities from the model

        - ``ips_mapping``: a dictionary mapping detected character tags to their associated

          IP (Intellectual Property) names in Danbooru-style format

        - ``ips_count``: a dictionary containing IP names and their occurrence counts based

          on detected characters

        - ``ips``: a list of IP names sorted by occurrence count (descending) and name

          (ascending), representing the most likely anime/game series in the image

        Default category thresholds are used if not specified. These vary by model and category

        but typically range from 0.35 to 0.5.

        >>> # Normalize for cosine similarity

        >>> import numpy as np

        >>> normalized_embedding = embedding / np.linalg.norm(embedding)

        >>>

        >>> # Get IP information for character identification

        >>> ips_mapping = get_pixai_tags('image.jpg', fmt='ips_mapping')

        >>> print("Character to IP mapping:", ips_mapping)

        >>> # Example output: {'misaka_mikoto': ['toaru_kagaku_no_railgun'], 'hu_tao_(genshin_impact)': ['genshin_impact']}

        >>>

        >>> # Get most likely anime/game series

        >>> top_ips = get_pixai_tags('image.jpg', fmt='ips')

        >>> print("Most likely series:", top_ips)

        >>> # Example output: ['genshin_impact', 'toaru_kagaku_no_railgun']

    """

    df_tags, d_ips = _open_tags(model_name=model_name)

    values = _raw_predict(image, model_name=model_name)