AssetAnalyzer

AssetAnalyzer#

class rapidtools.processing.AssetAnalyzer(model: BaseInferenceModel, prompt: str | Path = '', attribute_prefix: str | None = None, max_workers: int | None = None, batch_size: int = 4, max_images_per_asset: int | None = None, image_filter: Callable[[ImageAsset], bool] | None = None, cancel_event: Event | None = None, rate_limit: RateLimitPolicy | None = None, generation: GenerationConfig | None = None, mode: str | None = None)[source]#

Bases: object

Describe every asset in a collection with a vision-language model.

Parameters:

  • model (BaseInferenceModel) – Any rapidtools model wrapper (see rapidtools.models.load()).

  • prompt (str | Path) – Prompt text or path to a prompt file applied to every asset.

  • attribute_prefix (str | None) – Prefix of the attributes written to each asset. Defaults to the provider’s prefix ('gemini', 'qwen_vl', …).

  • max_workers (int | None) – Concurrent requests for hosted models. Defaults to 5 for APIs and 1 for local models (local inference is never threaded).

  • batch_size (int) – Assets per batch when the model offers run_inference_batch. Defaults to 4.

  • max_images_per_asset (int | None) – Cap on the images sent per asset; None sends them all.

  • image_filter (Callable[[ImageAsset], bool] | None) – Optional predicate selecting which images to analyze.

  • cancel_event (Event | None) – Cooperative cancellation flag checked between assets/batches.

  • rate_limit (RateLimitPolicy | None) – Back-off and retry policy for hosted models; None uses a 30-second cooldown with two retry passes for APIs and no cooldown for local models.

  • generation (GenerationConfig | None) – Per-call generation options (json_mode, temperature, …) forwarded to the model.

  • mode (str | None) – Force 'api' (threaded) or 'local' (sequential/batched) execution; inferred from the model when None.

Example

>>> from rapidtools.models import GenerationConfig, load
>>> from rapidtools.processing import AssetAnalyzer
>>> analyzer = AssetAnalyzer(
...     load('qwen_vl', model_id='Qwen/Qwen3.5-9B', load_in_4bit=True),
...     prompt='Return JSON with a damage_level key.',
...     batch_size=4,
...     generation=GenerationConfig(json_mode=True, temperature=0.0),
... )
>>> collection = analyzer(collection)
>>> collection.get('bldg_01').attributes['qwen_vl_damage_level']
'moderate'
ATTRIBUTE_PREFIX: str | None = None#

Default attribute prefix when neither the model nor the caller gives one.

property attribute_prefix: str#

Prefix of the attributes written to each asset.

property cooldown_duration: float#

Base cooldown in seconds (from rate_limit).