AssetAnalysisWorkflow

class rapidtools.gui.AssetAnalysisWorkflow(progress_callback: Callable[[str], None] | None = None)

Bases: object

Orchestrates asset detection and VLM inference on an aerial raster.

Parameters:

progress_callback (Callable[[str], None] | None) – Optional function receiving short human-readable status strings (e.g. 'Scanning for tree...'). The GUI uses this to update its status bar; scripts may pass print or leave it None.

Example

>>> wf = AssetAnalysisWorkflow()
>>> detection = wf.detect(DetectionSettings(
...     raster_path='eaton_patch_20250214.tiff',
...     output_dir='output',
...     assets=['building', 'tree'],
... ))
>>> inference = wf.analyze(detection.collection, InferenceSettings(
...     raster_path='eaton_patch_20250214.tiff',
...     output_dir='output',
...     prompt='Describe the damage to the outlined asset.',
...     api_key='...',
... ))

Methods

analyze	Crop the raster around every asset and run VLM inference on the crops.
assist	Run one prompt-assistant action with a rapidtools model backend.
cancel	Request cancellation.
detect	Detect one or more asset types in the raster with SAM 3.
discover_street	Discover objects along a Mapillary street survey.
download_basemap	Stitch a Bing or Google satellite basemap over a region into a GeoTIFF.
release_assistant	Drop the cached assistant model and free GPU memory if it was local.
reset	Clear any pending cancellation request.

cancel() → None

Request cancellation. Honoured at the next stage boundary.

reset() → None

Clear any pending cancellation request.

detect(settings: DetectionSettings) → DetectionResult

Detect one or more asset types in the raster with SAM 3.

Mirrors Part 1 of the generalized example notebook. When settings.detect_in_recon_imagery is False a Bing basemap is stitched over the raster extent and used for detection instead (useful when assets are destroyed in the post-disaster imagery).

Returns:

DetectionResult – The merged collection of every requested asset type plus the paths of the GeoJSON files written to disk.

analyze(collection: PhysicalAssetCollection, settings: InferenceSettings) → InferenceResult

Crop the raster around every asset and run VLM inference on the crops.

Mirrors Part 2 of the generalized example notebook.

Returns:

InferenceResult – The collection of assets that received at least one inferred attribute, and the GeoJSON path it was written to.

download_basemap(settings: RegionImagerySettings) → Path

Stitch a Bing or Google satellite basemap over a region into a GeoTIFF.

The region comes from the four WGS84 bounds or the extent of a GeoJSON file. The tile count is checked against settings.max_tiles before any download starts.

Parameters:

settings (RegionImagerySettings) – Provider, zoom and region.

Returns:

Path – The written GeoTIFF, named after the provider, zoom and bounds inside settings.output_dir.

Raises:

• ValueError – If the region covers more than max_tiles tiles.

• FileNotFoundError – If geojson_path does not exist.

discover_street(settings: StreetDetectionSettings) → StreetDetectionResult

Discover objects along a Mapillary street survey.

Wraps MapillaryFeatureExtractor: the requested classes are located in every image of the region from Mapillary’s segmentation metadata, the survey vehicle is removed, sightings are triangulated and clustered into one point asset per object. No pixels are downloaded; crops are made at inference time with the 'mapillary_objects' imagery source.

Parameters:

settings (StreetDetectionSettings) – Classes, token, region and localization options.

Returns:

StreetDetectionResult – The point assets and the GeoJSON they were written to (None when nothing was found).

Raises:

• FileNotFoundError – If raster_path is given but missing.

• WorkflowCancelled – If cancelled.

assist(settings: AssistSettings) → dict[str, Any]

Run one prompt-assistant action with a rapidtools model backend.

The model is loaded through rapidtools.models.load() and cached until a different backend or model is requested, or until detection or inference needs the GPU (local models are then released).

Parameters:

settings (AssistSettings) – The action and its inputs.

Returns:

dict[str, Any] – The run() result: spec, text or indicators.

Raises:

ValueError – If the model reply cannot be used.

release_assistant() → None

Drop the cached assistant model and free GPU memory if it was local.