xwhy.models.embeddings

Embedding module public API.

Registers and exposes available embedding implementations.

BaseEmbedding

Bases: ABC

Base class for all embedding implementations.

Source code in src/xwhy/models/embeddings/base.py

class BaseEmbedding(ABC):
    """Base class for all embedding implementations."""

    @property
    @abstractmethod
    def model(self) -> Any:  # noqa: ANN401
        """Read-only property to access the underlying raw embedding model.

        Raises:
            RuntimeError: If the model has not been loaded into memory yet.

        Returns:
            The loaded raw embedding model object.

        """
        pass

    @property
    @abstractmethod
    def processor(self) -> Any:  # noqa: ANN401
        """Read-only property to access the data processor/tokenizer.

        Returns:
            The associated processor object for the embedding model.

        """
        pass

    @abstractmethod
    def __call__(self, inputs: Any) -> Any:  # noqa: ANN401
        """Execute the forward pass to extract embeddings.

        Args:
            inputs: The preprocessed inputs ready for the model.

        Returns:
            The extracted embeddings/features.

        """
        pass

    @abstractmethod
    def load(self) -> Any:  # noqa: ANN401
        """Load embedding model into memory."""
        raise NotImplementedError

    @abstractmethod
    def encode(self, text: str) -> list[float]:
        """Encode text into vector representation."""
        raise NotImplementedError

model abstractmethod property

Read-only property to access the underlying raw embedding model.

Raises:

Type	Description
RuntimeError	If the model has not been loaded into memory yet.

Returns:

Type	Description
Any	The loaded raw embedding model object.

processor abstractmethod property

Read-only property to access the data processor/tokenizer.

Returns:

Type	Description
Any	The associated processor object for the embedding model.

__call__(inputs) abstractmethod

Execute the forward pass to extract embeddings.

Parameters:

Name	Type	Description	Default
inputs	Any	The preprocessed inputs ready for the model.	required

Returns:

Type	Description
Any	The extracted embeddings/features.

Source code in src/xwhy/models/embeddings/base.py

@abstractmethod
def __call__(self, inputs: Any) -> Any:  # noqa: ANN401
    """Execute the forward pass to extract embeddings.

    Args:
        inputs: The preprocessed inputs ready for the model.

    Returns:
        The extracted embeddings/features.

    """
    pass

load() abstractmethod

Load embedding model into memory.

Source code in src/xwhy/models/embeddings/base.py

@abstractmethod
def load(self) -> Any:  # noqa: ANN401
    """Load embedding model into memory."""
    raise NotImplementedError

encode(text) abstractmethod

Encode text into vector representation.

Source code in src/xwhy/models/embeddings/base.py

@abstractmethod
def encode(self, text: str) -> list[float]:
    """Encode text into vector representation."""
    raise NotImplementedError

EmbeddingFactory

Manage embedding model instantiation via a registry.

Source code in src/xwhy/models/embeddings/factory.py

class EmbeddingFactory:
    """Manage embedding model instantiation via a registry."""

    _registry: ClassVar[dict[EmbeddingType, Callable[..., BaseEmbedding]]] = {}

    @classmethod
    def register(
        cls, embedding: EmbeddingType, builder: Callable[..., BaseEmbedding]
    ) -> None:
        """Register a builder function for an embedding type.

        Args:
            embedding: The type of embedding to register.
            builder: A callable (function/lambda) that accepts keyword arguments
                and returns a BaseEmbedding instance.

        Raises:
            ValueError: If the embedding type is already registered.

        """
        if embedding in cls._registry:
            raise ValueError(f"Embedding already registered: {embedding}")
        cls._registry[embedding] = builder

    @classmethod
    def create(cls, embedding: EmbeddingType, **kwargs: object) -> BaseEmbedding:
        """Instantiate and configure an embedding model.

        Args:
            embedding: The type of embedding to create.
            **kwargs: Arbitrary keyword arguments passed to the builder function,
                such as 'settings' or 'model_name'.

        Returns:
            An instantiated BaseEmbedding object.

        Raises:
            ValueError: If the embedding type is not registered.

        """
        if embedding not in cls._registry:
            raise ValueError(f"Unsupported embedding: {embedding}")

        return cls._registry[embedding](**kwargs)

    @classmethod
    def clear(cls) -> None:
        """Reset registry to defaults."""
        cls._registry.clear()

register(embedding, builder) classmethod

Register a builder function for an embedding type.

Parameters:

Name	Type	Description	Default
embedding	EmbeddingType	The type of embedding to register.	required
builder	Callable[..., BaseEmbedding]	A callable (function/lambda) that accepts keyword arguments and returns a BaseEmbedding instance.	required

Raises:

Type	Description
ValueError	If the embedding type is already registered.

Source code in src/xwhy/models/embeddings/factory.py

@classmethod
def register(
    cls, embedding: EmbeddingType, builder: Callable[..., BaseEmbedding]
) -> None:
    """Register a builder function for an embedding type.

    Args:
        embedding: The type of embedding to register.
        builder: A callable (function/lambda) that accepts keyword arguments
            and returns a BaseEmbedding instance.

    Raises:
        ValueError: If the embedding type is already registered.

    """
    if embedding in cls._registry:
        raise ValueError(f"Embedding already registered: {embedding}")
    cls._registry[embedding] = builder

create(embedding, **kwargs) classmethod

Instantiate and configure an embedding model.

Parameters:

Name	Type	Description	Default
embedding	EmbeddingType	The type of embedding to create.	required
**kwargs	object	Arbitrary keyword arguments passed to the builder function, such as 'settings' or 'model_name'.	{}

Returns:

Type	Description
BaseEmbedding	An instantiated BaseEmbedding object.

Raises:

Type	Description
ValueError	If the embedding type is not registered.

Source code in src/xwhy/models/embeddings/factory.py

@classmethod
def create(cls, embedding: EmbeddingType, **kwargs: object) -> BaseEmbedding:
    """Instantiate and configure an embedding model.

    Args:
        embedding: The type of embedding to create.
        **kwargs: Arbitrary keyword arguments passed to the builder function,
            such as 'settings' or 'model_name'.

    Returns:
        An instantiated BaseEmbedding object.

    Raises:
        ValueError: If the embedding type is not registered.

    """
    if embedding not in cls._registry:
        raise ValueError(f"Unsupported embedding: {embedding}")

    return cls._registry[embedding](**kwargs)

clear() classmethod

Reset registry to defaults.

Source code in src/xwhy/models/embeddings/factory.py

@classmethod
def clear(cls) -> None:
    """Reset registry to defaults."""
    cls._registry.clear()

EmbeddingType

Bases: StrEnum

Supported embedding backends.

Source code in src/xwhy/models/embeddings/types.py

class EmbeddingType(StrEnum):
    """Supported embedding backends."""

    WORD2VEC = "word2vec"
    GLOVE = "glove"
    PARAGRAM_SL = "paragram_sl"
    PARAGRAM_WS = "paragram_ws"
    DINOV2 = "dinov2"

    # Future:
    # SENTENCE_TRANSFORMER = "sentence_transformer"
    # CLIP = "clip"
    # BGE = "bge"

    @classmethod
    def from_str(cls, value: str | EmbeddingType) -> EmbeddingType:
        """Safely convert a string or enum instance to EmbeddingType."""
        try:
            return cls(value)
        except ValueError as err:
            valid_options = ", ".join([item.value for item in cls])
            raise ValueError(
                f"'{value}' is not a valid EmbeddingType. "
                f"Supported options are: [{valid_options}]"
            ) from err

    @property
    def is_image_embedding(self) -> bool:
        """Check if this is an image-based embedding."""
        return self in {EmbeddingType.DINOV2}

    @property
    def is_text_embedding(self) -> bool:
        """Check if this is a text-based embedding."""
        return self in {
            EmbeddingType.WORD2VEC,
            EmbeddingType.GLOVE,
            EmbeddingType.PARAGRAM_SL,
            EmbeddingType.PARAGRAM_WS,
        }

is_image_embedding property

Check if this is an image-based embedding.

is_text_embedding property

Check if this is a text-based embedding.

from_str(value) classmethod

Safely convert a string or enum instance to EmbeddingType.

Source code in src/xwhy/models/embeddings/types.py

@classmethod
def from_str(cls, value: str | EmbeddingType) -> EmbeddingType:
    """Safely convert a string or enum instance to EmbeddingType."""
    try:
        return cls(value)
    except ValueError as err:
        valid_options = ", ".join([item.value for item in cls])
        raise ValueError(
            f"'{value}' is not a valid EmbeddingType. "
            f"Supported options are: [{valid_options}]"
        ) from err

Word2VecEmbedding

Bases: BaseEmbedding

Word2Vec embedding backend.

Source code in src/xwhy/models/embeddings/word2vec.py

class Word2VecEmbedding(BaseEmbedding):
    """Word2Vec embedding backend."""

    _MODEL_FILE_MAP: ClassVar[dict[str, dict[str, Any]]] = {
        "word2vec-google-news-300": {
            "file": "GoogleNews-vectors-negative300.bin",
            "binary": True,
            "gensim": True,
            "no_header": False,
            "google_id": "1vAjPzr5R1RQiuh9NOgHVFGhBmCuYWnJU",
        },
        "glove.840B.300d": {
            "file": "glove.840B.300d.txt",
            "binary": False,
            "gensim": True,
            "no_header": True,
            "google_id": "19cJAKkgrYAiT1gU-OnWTN6GaGcd7pLZI",
        },
        "paragram_300_sl999": {
            "file": "paragram_300_sl999.txt",
            "binary": False,
            "gensim": True,
            "no_header": True,
            "google_id": "1c-16FP0jvaJeyaM8JcqqKPdoVw7uqVkK",
        },
        "paragram-300-WS353": {
            "file": "paragram_300_ws353.txt",
            "binary": False,
            "gensim": True,
            "no_header": True,
            "google_id": "1bBLz6F6MJ_qx9xnSZUgI8W0QZhJYgBdu",
        },
    }

    def __init__(
        self,
        *,
        settings: Settings,
        model_name: str = "word2vec-google-news-300",
        force_download: bool = False,
        **kwargs: Any,  # noqa: ANN401
    ) -> None:
        """Initialize Word2Vec embedding backend."""
        self._settings = settings
        self._model_name = model_name
        self._force_download = force_download
        self._model: KeyedVectors | None = None

    @property
    def model(self) -> Any:  # noqa: ANN401
        """Read-only property to access the underlying Word2Vec model."""
        if self._model is None:
            raise RuntimeError("Word2Vec model is not loaded. Call .load() first.")
        return self._model

    @property
    def processor(self) -> Any:  # noqa: ANN401
        """Read-only dummy processor property for interface compliance.

        Returns:
            None, as Word2Vec does not require an external HuggingFace processor.

        """
        return None

    def __call__(
        self,
        inputs: Any,  # noqa: ANN401
        **kwargs: Any,  # noqa: ANN401
    ) -> np.ndarray | list[np.ndarray]:
        """Encode inputs into word embeddings.

        Args:
            inputs: Source text string or a sequence (tuple/list) of strings.
            **kwargs: Additional keyword arguments (kept for interface compatibility).

        Returns:
            np.ndarray | list[np.ndarray]: A 2D array of embeddings for a single string,
                or a list of 2D arrays if a sequence of strings is provided.

        """
        if isinstance(inputs, (list, tuple)):
            return [self.encode(text=str(item)) for item in inputs]

        return self.encode(text=str(inputs))

    def load(self) -> KeyedVectors:
        """Load embedding model with caching strategy."""
        if self._model is not None:
            return self._model

        cache_dir = self._get_cache_dir()
        cache_dir.mkdir(parents=True, exist_ok=True)

        if self._model_name not in self._MODEL_FILE_MAP:
            raise ValueError(f"Unsupported model: {self._model_name}")

        model_info = self._MODEL_FILE_MAP[self._model_name]
        original_model_path = cache_dir / model_info["file"]

        # 1. cache
        bin_cache_path = original_model_path.with_suffix(".bin")

        if bin_cache_path.exists() and not self._force_download:
            logger.debug(
                f"Loading embedding model from fast binary cache: {bin_cache_path}"
            )
            self._model = KeyedVectors.load_word2vec_format(
                str(bin_cache_path),
                binary=True,
            )
            return self._model

        # 2. gdown direct download option
        gdown_model = self._try_gdown_download(
            model_info=model_info,
            bin_cache_path=bin_cache_path,
        )
        if gdown_model is not None:
            return gdown_model

        # 3. gensim
        if model_info["gensim"]:
            logger.debug(f"Loading embedding model via gensim: {self._model_name}")
            try:
                model = api.load(self._model_name)
                logger.info(f"Saving to fast binary cache: {bin_cache_path}")
                model.save_word2vec_format(str(bin_cache_path), binary=True)
                self._model = model
                return model

            except Exception:
                pass

        # 4. fallback download for specific models
        if self._model_name == "word2vec-google-news-300":
            return self._download_google_news(cache_dir, bin_cache_path)
        elif self._model_name == "glove.840B.300d":
            return self._download_glove(cache_dir, bin_cache_path, original_model_path)
        elif self._model_name in ("paragram_300_sl999", "paragram-300-WS353"):
            return self._download_paragram(
                cache_dir, bin_cache_path, original_model_path
            )

        raise RuntimeError(f"Failed to load embedding model: {self._model_name}")

    def encode(self, text: str) -> np.ndarray:  # type: ignore[override]
        """Extract a 2D array of word embeddings for a given text.

        Args:
            text: The raw input text string to be embedded.

        Returns:
            np.ndarray: A 2D numpy array of shape (N, D) where N is the number of
                in-vocabulary words and D is the embedding dimension. Returns an
                empty array of shape (0, D) if no valid words are found.

        """
        model = self.load()

        cleaned = clean_text(text=text)

        words = [word for word in cleaned.split() if word in model]

        if not words:
            # Return an empty 2D array with the correct feature dimension
            return np.empty((0, model.vector_size), dtype=float)

        vectors = [model[word] for word in words]
        return np.array(vectors, dtype=float)

    def _get_cache_dir(self) -> Path:
        cache_dir = self._settings.embedding_cache_dir
        return cache_dir if cache_dir else Path(Path.home() / ".cache/xwhy/embeddings")

    def _try_gdown_download(
        self,
        model_info: dict[str, Any],
        bin_cache_path: Path,
    ) -> KeyedVectors | None:
        """Attempt to download pre-converted binary model from Google Drive."""
        google_id: str | None = model_info.get("google_id")

        if not google_id or self._force_download:
            return None

        logger.debug(f"Attempting binary download via gdown for ID: {google_id}")

        try:
            # Download directly to bin_cache_path since the GDrive file is already .bin
            gdown.download(id=google_id, output=str(bin_cache_path), quiet=False)  # type: ignore[attr-defined]

            # All files on GDrive are pre-converted binary format. Force binary=True.
            model = KeyedVectors.load_word2vec_format(
                str(bin_cache_path),
                binary=True,
                no_header=False,
            )

            self._model = model
            return self._model

        except Exception as err:
            logger.debug(f"gdown route failed for {self._model_name}: {err}")
            if bin_cache_path.exists():
                bin_cache_path.unlink()
            return None

    def _download_google_news(
        self,
        cache_dir: Path,
        bin_cache_path: Path,
    ) -> KeyedVectors:
        """Download GoogleNews model."""
        gz_path = cache_dir / "GoogleNews-vectors-negative300.bin.gz"
        url = (
            "https://public.ukp.informatik.tu-darmstadt.de/"
            "reimers/wordembeddings/GoogleNews-vectors-negative300.bin.gz"
        )

        try:
            if not gz_path.exists():
                self._download_file(url, gz_path)

            self._extract_gzip(gz_path, bin_cache_path)

            self._model = KeyedVectors.load_word2vec_format(
                str(bin_cache_path),
                binary=True,
            )

            return self._model

        except Exception as error:
            raise RuntimeError("Failed to load GoogleNews model") from error

    def _download_glove(
        self, cache_dir: Path, bin_path: Path, txt_path: Path
    ) -> KeyedVectors:
        """Download, clean, and convert GloVe model to binary."""
        zip_path = cache_dir / "glove.840B.300d.zip"
        url = "http://nlp.stanford.edu/data/glove.840B.300d.zip"

        try:
            if not zip_path.exists():
                self._download_file(url, zip_path)

            if not txt_path.exists():
                logger.info("Extracting GloVe zip file...")
                with zipfile.ZipFile(zip_path, "r") as zip_ref:
                    zip_ref.extractall(cache_dir)

            cleaned_path = txt_path.with_name(txt_path.name + ".cleaned")
            if not cleaned_path.exists():
                seen = set()
                logger.info("De-duplicating GloVe file...")
                with (
                    open(txt_path, encoding="utf-8", errors="ignore") as fin,
                    open(cleaned_path, "w", encoding="utf-8") as fout,
                ):
                    for line in tqdm(fin, desc="Cleaning GloVe"):
                        word = line.split(maxsplit=1)[0]
                        if word not in seen:
                            fout.write(line)
                            seen.add(word)

            logger.info("Loading cleaned GloVe text...")
            model = KeyedVectors.load_word2vec_format(
                str(cleaned_path), binary=False, no_header=True
            )

            logger.info("Converting GloVe to fast binary format...")
            model.save_word2vec_format(str(bin_path), binary=True)

            for path_to_remove in [cleaned_path, txt_path, zip_path]:
                if path_to_remove.exists():
                    path_to_remove.unlink()

            self._model = model
            return self._model
        except Exception as error:
            raise RuntimeError("Failed to load GloVe model") from error

    def _download_paragram(
        self, cache_dir: Path, bin_path: Path, txt_path: Path
    ) -> KeyedVectors:
        """Download and convert Paragram models to binary."""
        gdrive_ids = {
            "paragram_300_sl999": "0B9w48e1rj-MOck1fRGxaZW1LU2M",
            "paragram-300-WS353": "0B9w48e1rj-MOLVdZRzFfTlNsem8",
        }
        file_id = gdrive_ids[self._model_name]

        try:
            if not txt_path.exists():
                logger.info(f"Downloading {self._model_name} via gdown...")

                temp_path = txt_path.with_suffix(".tmp")
                gdown.download(id=file_id, output=str(temp_path), quiet=False)  # type: ignore[attr-defined]

                if zipfile.is_zipfile(str(temp_path)):
                    logger.info("Extracting zip archive...")
                    with zipfile.ZipFile(str(temp_path), "r") as zf:
                        # Find the text file name inside the zip archive
                        txt_filename = next(
                            (name for name in zf.namelist() if name.endswith(".txt")),
                            None,
                        )
                        if not txt_filename:
                            raise FileNotFoundError("No .txt file found inside zip.")

                        # Manually extract with Python to bypass CRC errors
                        with open(txt_path, "wb") as f_out:
                            try:
                                with zf.open(txt_filename) as f_in:
                                    shutil.copyfileobj(f_in, f_out)
                            except zipfile.BadZipFile:
                                logger.warning("Ignored CRC error.")
                    temp_path.unlink()
                else:
                    temp_path.rename(txt_path)

            logger.info("Sanitizing text file and adding header...")
            clean_txt_path = txt_path.with_suffix(".clean.txt")
            expected_dim = 300

            # Step 1: Count valid lines (lines with exactly 301
            # parts: 1 word + 300 numbers)
            valid_lines = 0
            with open(txt_path, encoding="utf-8", errors="ignore") as f:
                for line in f:
                    if len(line.rstrip("\n").split(" ")) == expected_dim + 1:
                        valid_lines += 1

            # Step 2: Rewrite the file with only valid lines
            # + add standard word2vec header
            with (
                open(txt_path, encoding="utf-8", errors="ignore") as f_in,
                open(clean_txt_path, "w", encoding="utf-8") as f_out,
            ):
                f_out.write(f"{valid_lines} {expected_dim}\n")
                for line in f_in:
                    if len(line.rstrip("\n").split(" ")) == expected_dim + 1:
                        f_out.write(line)

            logger.info(f"Loading {self._model_name} text into Gensim...")
            # Since we added the header to the cleaned file
            # ourselves, set no_header=False
            model = KeyedVectors.load_word2vec_format(
                str(clean_txt_path),
                binary=False,
                no_header=False,
                unicode_errors="ignore",
            )

            logger.info(f"Converting {self._model_name} to binary...")
            model.save_word2vec_format(str(bin_path), binary=True)

            # Clean up both text files (original and cleaned) to free up disk space
            if txt_path.exists():
                txt_path.unlink()
            if clean_txt_path.exists():
                clean_txt_path.unlink()

            self._model = model
            return self._model
        except Exception as error:
            raise RuntimeError(f"Failed to load {self._model_name}") from error

    @staticmethod
    def _download_file(url: str, path: Path) -> None:
        """Download a file using streaming requests."""
        path.parent.mkdir(parents=True, exist_ok=True)
        try:
            logger.debug(f"Attempting direct download from {url}...")
            response = requests.get(url, stream=True, timeout=600)
            response.raise_for_status()
            total_size = int(response.headers.get("content-length", 0))

            with (
                path.open("wb") as file,
                tqdm(
                    desc=path.name,
                    total=total_size,
                    unit="iB",
                    unit_scale=True,
                    unit_divisor=1024,
                ) as bar,
            ):
                for chunk in response.iter_content(chunk_size=8192):
                    if chunk:
                        size = file.write(chunk)
                        bar.update(size)

            if path.stat().st_size < 100:
                raise OSError("Downloaded file too small")

        except Exception:
            if path.exists():
                path.unlink()

            raise

    @staticmethod
    def _extract_gzip(src: Path, dst: Path) -> None:
        """Extract gzip file."""
        logger.debug("Extracting %s...", src)
        with gzip.open(src, "rb") as f_in, dst.open("wb") as f_out:
            shutil.copyfileobj(f_in, f_out)

model property

Read-only property to access the underlying Word2Vec model.

processor property

Read-only dummy processor property for interface compliance.

Returns:

Type	Description
Any	None, as Word2Vec does not require an external HuggingFace processor.

__init__(*, settings, model_name='word2vec-google-news-300', force_download=False, **kwargs)

Initialize Word2Vec embedding backend.

Source code in src/xwhy/models/embeddings/word2vec.py

def __init__(
    self,
    *,
    settings: Settings,
    model_name: str = "word2vec-google-news-300",
    force_download: bool = False,
    **kwargs: Any,  # noqa: ANN401
) -> None:
    """Initialize Word2Vec embedding backend."""
    self._settings = settings
    self._model_name = model_name
    self._force_download = force_download
    self._model: KeyedVectors | None = None

__call__(inputs, **kwargs)

Encode inputs into word embeddings.

Parameters:

Name	Type	Description	Default
inputs	Any	Source text string or a sequence (tuple/list) of strings.	required
**kwargs	Any	Additional keyword arguments (kept for interface compatibility).	{}

Returns:

Type	Description
ndarray | list[ndarray]	np.ndarray | list[np.ndarray]: A 2D array of embeddings for a single string, or a list of 2D arrays if a sequence of strings is provided.

Source code in src/xwhy/models/embeddings/word2vec.py

def __call__(
    self,
    inputs: Any,  # noqa: ANN401
    **kwargs: Any,  # noqa: ANN401
) -> np.ndarray | list[np.ndarray]:
    """Encode inputs into word embeddings.

    Args:
        inputs: Source text string or a sequence (tuple/list) of strings.
        **kwargs: Additional keyword arguments (kept for interface compatibility).

    Returns:
        np.ndarray | list[np.ndarray]: A 2D array of embeddings for a single string,
            or a list of 2D arrays if a sequence of strings is provided.

    """
    if isinstance(inputs, (list, tuple)):
        return [self.encode(text=str(item)) for item in inputs]

    return self.encode(text=str(inputs))

load()

Load embedding model with caching strategy.

Source code in src/xwhy/models/embeddings/word2vec.py

def load(self) -> KeyedVectors:
    """Load embedding model with caching strategy."""
    if self._model is not None:
        return self._model

    cache_dir = self._get_cache_dir()
    cache_dir.mkdir(parents=True, exist_ok=True)

    if self._model_name not in self._MODEL_FILE_MAP:
        raise ValueError(f"Unsupported model: {self._model_name}")

    model_info = self._MODEL_FILE_MAP[self._model_name]
    original_model_path = cache_dir / model_info["file"]

    # 1. cache
    bin_cache_path = original_model_path.with_suffix(".bin")

    if bin_cache_path.exists() and not self._force_download:
        logger.debug(
            f"Loading embedding model from fast binary cache: {bin_cache_path}"
        )
        self._model = KeyedVectors.load_word2vec_format(
            str(bin_cache_path),
            binary=True,
        )
        return self._model

    # 2. gdown direct download option
    gdown_model = self._try_gdown_download(
        model_info=model_info,
        bin_cache_path=bin_cache_path,
    )
    if gdown_model is not None:
        return gdown_model

    # 3. gensim
    if model_info["gensim"]:
        logger.debug(f"Loading embedding model via gensim: {self._model_name}")
        try:
            model = api.load(self._model_name)
            logger.info(f"Saving to fast binary cache: {bin_cache_path}")
            model.save_word2vec_format(str(bin_cache_path), binary=True)
            self._model = model
            return model

        except Exception:
            pass

    # 4. fallback download for specific models
    if self._model_name == "word2vec-google-news-300":
        return self._download_google_news(cache_dir, bin_cache_path)
    elif self._model_name == "glove.840B.300d":
        return self._download_glove(cache_dir, bin_cache_path, original_model_path)
    elif self._model_name in ("paragram_300_sl999", "paragram-300-WS353"):
        return self._download_paragram(
            cache_dir, bin_cache_path, original_model_path
        )

    raise RuntimeError(f"Failed to load embedding model: {self._model_name}")

encode(text)

Extract a 2D array of word embeddings for a given text.

Parameters:

Name	Type	Description	Default
text	str	The raw input text string to be embedded.	required

Returns:

Type	Description
ndarray	np.ndarray: A 2D numpy array of shape (N, D) where N is the number of in-vocabulary words and D is the embedding dimension. Returns an empty array of shape (0, D) if no valid words are found.

Source code in src/xwhy/models/embeddings/word2vec.py

def encode(self, text: str) -> np.ndarray:  # type: ignore[override]
    """Extract a 2D array of word embeddings for a given text.

    Args:
        text: The raw input text string to be embedded.

    Returns:
        np.ndarray: A 2D numpy array of shape (N, D) where N is the number of
            in-vocabulary words and D is the embedding dimension. Returns an
            empty array of shape (0, D) if no valid words are found.

    """
    model = self.load()

    cleaned = clean_text(text=text)

    words = [word for word in cleaned.split() if word in model]

    if not words:
        # Return an empty 2D array with the correct feature dimension
        return np.empty((0, model.vector_size), dtype=float)

    vectors = [model[word] for word in words]
    return np.array(vectors, dtype=float)