cjm-transcription-plugin-gemini 0.0.21

Google Gemini API plugin for the cjm-transcription-plugin-system library - provides speech-to-text transcription with configurable model selection and parameter control.

cjm-transcription-plugin-gemini

Install

pip install cjm_transcription_plugin_gemini

Project Structure

nbs/
├── meta.ipynb   # Metadata introspection for the Gemini plugin used by cjm-ctl to generate the registration manifest.
└── plugin.ipynb # Plugin implementation for Google Gemini API transcription

Total: 2 notebooks across 1 directory

Module Dependencies

graph LR
    meta[meta<br/>Metadata]
    plugin[plugin<br/>Gemini Plugin]

    plugin --> meta

1 cross-module dependencies detected

CLI Reference

No CLI commands found in this project.

Module Overview

Detailed documentation for each module in the project:

Metadata (meta.ipynb)

Metadata introspection for the Gemini plugin used by cjm-ctl to generate the registration manifest.

Import

from cjm_transcription_plugin_gemini.meta import (
    get_plugin_metadata
)

Functions

def get_plugin_metadata() -> Dict[str, Any]: # Plugin metadata for manifest generation
    """Return metadata required to register this plugin with the PluginManager."""
    # Fallback base path (current behavior for backward compatibility)
    base_path = os.path.dirname(os.path.dirname(sys.executable))
    
    # Use CJM config if available, else fallback to env-relative paths
    cjm_data_dir = os.environ.get("CJM_DATA_DIR")
    
    # Plugin data directory
    plugin_name = "cjm-transcription-plugin-gemini"
    if cjm_data_dir
    "Return metadata required to register this plugin with the PluginManager."

Gemini Plugin (plugin.ipynb)

Plugin implementation for Google Gemini API transcription

Import

from cjm_transcription_plugin_gemini.plugin import (
    GeminiPluginConfig,
    GeminiPlugin
)

Functions

@patch
def _refresh_available_models(
    self:GeminiPlugin
) -> List[str]:  # Live audio-capable model names (empty if no client / API unreachable)
    """
    Fetch + filter audio-capable models from the LIVE Gemini API (CR-11).
    
    Returns [] when there's no client or the API can't be reached — there is no
    static fallback list (a stale model list is worse than none; a baked default
    rots as Google retires/adds models). Side effect: repopulates
    self.model_token_limits (model name -> output_token_limit) for the live set.
    """

@patch
def _update_max_tokens_for_model(
    self:GeminiPlugin,
    model_name: str  # Model name to update tokens for
) -> None
    "Clamp max_output_tokens to the model's live output_token_limit when known."

@patch
def update_config(
    self:GeminiPlugin,
    config: Union[Dict[str, Any], GeminiPluginConfig]  # New configuration values
) -> None
    "Update plugin configuration, adjusting max_tokens if model changes."

@patch
def _prepare_audio(
    self:GeminiPlugin,
    audio: Union[str, Path]  # Path to a decodable audio file
) -> Tuple[Path, bool]:  # Tuple of (processed audio path, whether temp file was created)
    """
    Prepare audio file for upload.
    
    The caller provides a decodable audio file path; in-memory preparation is no
    longer a plugin responsibility. Optional downsampling (a relocation candidate
    that belongs in an upstream ffmpeg pipeline step) is retained for now.
    """

@patch
def _upload_audio_file(
    self:GeminiPlugin,
    audio_path: Path  # Path to audio file to upload
) -> Any:  # Uploaded file object
    "Upload audio file to Gemini API."

@patch
def _delete_uploaded_file(
    self:GeminiPlugin,
    file_name: str  # Name of file to delete
) -> None
    "Delete an uploaded file from Gemini API."

@patch
def get_available_models(
    self:GeminiPlugin
) -> List[str]:  # Live audio-capable model names (empty if no API key / API unreachable)
    """
    Get the live list of audio-capable models (CR-11).
    
    Lazily ensures the client from the injected API key, then refreshes from the
    API. Returns [] when no key is available — no static fallback.
    """

@patch
def get_model_info(
    self:GeminiPlugin,
    model_name: Optional[str] = None  # Model name to get info for, defaults to current model
) -> Dict[str, Any]:  # Dict with model information
    "Get information about a specific model including its token limit."

@patch
def supports_streaming(
    self:GeminiPlugin
) -> bool:  # True if streaming is supported
    "Check if this plugin supports streaming transcription."

@patch
def execute_stream(
    self:GeminiPlugin,
    audio: Union[str, Path],  # Audio data object or path to audio file
    **kwargs  # Additional arguments to override config
) -> Generator[str, None, TranscriptionResult]:  # Yields text chunks, returns final result
    "Stream transcription results chunk by chunk."

Classes

@dataclass
class GeminiPluginConfig:
    """
    Configuration for Gemini transcription plugin.
    
    CR-11/CR-12 notes:
      - `model` has NO static enum and NO default. Its valid domain is the live
        Gemini model list, surfaced at runtime via `get_config_options()` — a
        baked-in enum/default goes stale (models are retired/added server-side),
        and if the live list can't be fetched the API is unusable anyway, so a
        default selection is worthless. `model` is None until the operator picks
        one from the live list.
      - The API key is NOT a config field. It is a secret declared in
        `GeminiPlugin.WORKER_ENV` and resolved from the SecretStore into the
        worker env at spawn (see CR-12).
    """
    
    model: Optional[str] = field(...)
    prompt: str = field(...)
    temperature: float = field(...)
    top_p: float = field(...)
    max_output_tokens: int = field(...)
    seed: Optional[int] = field(...)
    response_mime_type: str = field(...)
    downsample_audio: bool = field(...)
    downsample_rate: int = field(...)
    downsample_channels: int = field(...)
    safety_settings: str = field(...)
    model_filter: List[str] = field(...)
    use_file_upload: bool = field(...)
    use_streaming: bool = field(...)
    delete_uploaded_files: bool = field(...)

class GeminiPlugin:
    def __init__(self):
        """Initialize the Gemini plugin with default configuration."""
        self.logger = logging.getLogger(f"{__name__}.{type(self).__name__}")
        self.config: GeminiPluginConfig = None
    "Google Gemini API transcription plugin."
    
    def __init__(self):
            """Initialize the Gemini plugin with default configuration."""
            self.logger = logging.getLogger(f"{__name__}.{type(self).__name__}")
            self.config: GeminiPluginConfig = None
        "Initialize the Gemini plugin with default configuration."
    
    def name(self) -> str: # Plugin name identifier
            """Return the plugin name identifier."""
            return "gemini"
        
        @property
        def version(self) -> str: # Plugin version string
        "Return the plugin name identifier."
    
    def version(self) -> str: # Plugin version string
            """Return the plugin version string."""
            return "1.0.0"
        
        @property
        def supported_formats(self) -> List[str]: # List of supported audio formats
        "Return the plugin version string."
    
    def supported_formats(self) -> List[str]: # List of supported audio formats
            """Return list of supported audio file formats."""
            return ["wav", "mp3", "aiff", "aac", "ogg", "flac"]
    
        def get_current_config(self) -> Dict[str, Any]: # Current configuration as dictionary
        "Return list of supported audio file formats."
    
    def get_current_config(self) -> Dict[str, Any]: # Current configuration as dictionary
            """Return current configuration state."""
            if not self.config
        "Return current configuration state."
    
    def get_config_schema(self) -> Dict[str, Any]: # JSON Schema for configuration
            """Return JSON Schema for UI generation."""
            return dataclass_to_jsonschema(GeminiPluginConfig)
    
        @staticmethod
        def get_config_dataclass() -> GeminiPluginConfig: # Configuration dataclass
        "Return JSON Schema for UI generation."
    
    def get_config_dataclass() -> GeminiPluginConfig: # Configuration dataclass
            """Return dataclass describing the plugin's configuration options."""
            return GeminiPluginConfig
        
        def initialize(
            self,
            config: Optional[Any] = None # Configuration dataclass, dict, or None
        ) -> None
        "Return dataclass describing the plugin's configuration options."
    
    def initialize(
            self,
            config: Optional[Any] = None # Configuration dataclass, dict, or None
        ) -> None
        "First-time setup (CR-4): apply config + set up storage.

Does NOT eagerly create the API client or fetch the model list — the
client is created lazily (see `_ensure_client`) once the GEMINI_API_KEY
is present in the worker env. This lets the plugin LOAD without a key
(so a config UI can collect one post-load), and means the live model
list (`get_config_options`) only populates once the key is set + the
worker has respawned with it injected."
    
    def execute(
            self,
            audio: Union[str, Path], # Audio data object or path to audio file
            **kwargs # Additional arguments to override config
        ) -> TranscriptionResult: # Transcription result object
        "Transcribe audio using Gemini."
    
    def is_available(self) -> bool: # True if the Gemini API is available
            """Check if Gemini API is available."""
            return GEMINI_AVAILABLE
    
        def get_config_options(self) -> Dict[str, "FieldOptions"]
        "Check if Gemini API is available."
    
    def get_config_options(self) -> Dict[str, "FieldOptions"]:
            """CR-11: live config option domains, keyed by config field name.
    
            Surfaces the live Gemini model list for `model` (with per-model output
            token limits as option metadata + a derived `max_output_tokens` ceiling
            constraint for the currently-selected model). Returns {} when no client
            is available (no API key yet) or the API can't be reached — there is no
            useful static fallback (a stale model list is worse than none), so the
            UI shows 'set an API key first' rather than a frozen enum.
            """
            self._ensure_client()
            if self.client is None
        "CR-11: live config option domains, keyed by config field name.

Surfaces the live Gemini model list for `model` (with per-model output
token limits as option metadata + a derived `max_output_tokens` ceiling
constraint for the currently-selected model). Returns {} when no client
is available (no API key yet) or the API can't be reached — there is no
useful static fallback (a stale model list is worse than none), so the
UI shows 'set an API key first' rather than a frozen enum."
    
    def prefetch(self) -> None:
            """CR-4: eagerly create the client + warm the live model list.
    
            Optional pre-warm so the first get_config_options / execute is fast.
            No-op (client stays None) when no API key is injected yet.
            """
            self._ensure_client()
            if self.client is not None
        "CR-4: eagerly create the client + warm the live model list.

Optional pre-warm so the first get_config_options / execute is fast.
No-op (client stays None) when no API key is injected yet."
    
    def on_disable(self) -> None:
            """CR-2: release the API client when the operator disables the plugin.
    
            Lightweight for an API plugin (no GPU/model memory), but drops the client
            so a re-enable / next execute lazily re-creates it from the current env.
            """
            self.client = None
    
        def cleanup(
            self
        ) -> None
        "CR-2: release the API client when the operator disables the plugin.

Lightweight for an API plugin (no GPU/model memory), but drops the client
so a re-enable / next execute lazily re-creates it from the current env."
    
    def cleanup(
            self
        ) -> None
        "Clean up resources."