Search URLs

Wheel encodes the full search state in the URL hash. Copying the browser URL captures every active search mode, filter, and parameter, which makes it possible to bookmark a slice, share it with a colleague, or paste it into a tracker as a reproducible reference. The same encoding is also useful when constructing a link to Wheel programmatically, for example from a notebook, a script, or the Agent Interface.

This page documents the URL structure and lists the parameters that correspond to each search mode in Search. Every parameter described here can also be set through the UI controls; the URL is just the serialised form of the same configuration.

URL Structure

A Wheel search URL has the form:

http://your-wheel-server/#&page=0&<key>=<value>&<key>=<value>...

The fragment starts with #& (the leading & is intentional and is part of how Wheel parses the hash) and is followed by key=value pairs joined by &. page=0 selects the first page of results and is usually included when sharing a URL.

Encoding rules:

• Values are URL-encoded. Spaces are encoded as %20 (not +), and symbols such as > or & inside a value are percent-encoded.

• Multi-value filters are joined with the literal token || (encoded as %7C%7C). The exceptions are cluster_ids and wm_angle_range, which use a single comma.

• Boolean values are serialised as the strings true and false.

• Omitted parameters fall back to their defaults. project_source defaults to Alpamayo and is the only field that has a meaningful implicit default; everything else is simply unset when absent.

Common Parameters

These parameters apply across all search modes and define which project, which datasets, and how results are ordered.

Parameter	Description	Notes
page	Result page index, starting at 0.	Integer.
project_source	One or more project names whose annotations and labels are loaded.	|| separated. Defaults to Alpamayo.
data_source	One or more datasets to restrict the search to. Examples currently documented include MADS, MADS-1M, AV V1 train, AV V1 validation, AV V2 train, AV V2 validation, Nexar, ncore-lidar-model-static-full, and vipe-dynpose100kpp. The full list is deployment-specific and is shown in the Data Source dropdown of the search panel.	|| separated. See Metadata Search.
data_source_mode	Combination rule when multiple data sources are selected: any (clip belongs to at least one) or all (clip belongs to every listed source).	Defaults to any.
rank_mode	Ordering strategy when multiple ranked search modes are active. priority (default) uses the fixed priority order described in Search. rrf merges the active ranked modes with reciprocal rank fusion, giving each mode comparable influence on the final order.	Defaults to priority.

Caption Search

Parameters that drive Caption Search and Caption Embedding Search.

Parameter	Description
search	Caption keyword query. Supports quoted phrases and the boolean operators AND, OR, NOT, as described in Caption Search.
caption_extra_queries	||-separated list of rewritten variants generated by the Rewrite action. Each variant is OR-ed with the original search query.
query_rewrite	true enables LLM-driven query rewriting at search time.
caption_embed_search	Free-form text used as the query for caption embedding search (Qwen3 embeddings).
caption_embed_extra_queries	||-separated list of additional caption embedding queries merged into the same ranked retrieval.

Semantic and Visual Search

Parameters that drive Semantic Search and Visual Search.

Parameter	Description
semantic_search_text	Text query for semantic text-to-video retrieval (CosmosEmbed1).
semantic_search_clipid	Reference clip ID for semantic video-to-video retrieval.
semantic_extra_queries	||-separated list of additional semantic queries merged with semantic_search_text.
visual_search_text	Text query for frame-level visual search (CLIP).
visual_search_image_id	Image ID used as the visual query for image-conditioned search.
visual_extra_queries	||-separated list of additional visual queries merged with visual_search_text.

Trajectory Search

Parameters that drive Trajectory Search. Both shape search and pattern search require ego trajectory data, so with_ego_data=true is implied.

Parameter	Description
trajectory_pattern	Named pattern (e.g. hard_braking, high_curvature, stop_go) or a custom motion expression over the variables documented in Trajectory Search.
trajectory_shape_clipid	Reference clip ID whose ego path is used as the shape target.
trajectory_shape_start_t	Start time, in seconds, of the sub-segment of the reference clip used for shape comparison.
trajectory_shape_end_t	End time, in seconds, of the sub-segment of the reference clip.
search_speed	Speed-based pattern shortcut.
with_ego_data	true restricts results to clips with ego trajectory data attached.

Perception-based Search

Parameters that drive Perception-based Search. Note that the URL parameter names use the wm_ (world model) prefix.

Parameter	Description
wm_class_name	Object class to filter on, from VEHICLE_CAR, VEHICLE_TRUCK, VEHICLE_BUS, BIKE_WITH_RIDER, BIKE_TRICYCLE, PEDESTRIAN_UNKNOWN.
wm_min_count / wm_max_count	Lower and upper bounds on the total detection count of the selected class across the clip.
wm_max_dist	Maximum distance, in metres, between the object and the ego vehicle for the range filter.
wm_min_time	Minimum continuous duration, in seconds, that the object must remain in the configured spatial window.
wm_angle_range	Comma-separated list of sectors from FRONT, FRONT_RIGHT, BACK_RIGHT, BACK, BACK_LEFT, FRONT_LEFT. Note the comma delimiter, which differs from the || used elsewhere.

Classifier Search

Parameters that drive Classifier Search.

Parameter	Description
classifier_run_id	Identifier of the trained classifier run to apply.
probability_expression	Expression on the classifier probability p, for example p > 0.95. Comparison and logical operators are supported.

Cluster Search

Parameters that drive Cluster Search.

Parameter	Description
cluster_run_id	Identifier of the clustering run to browse.
cluster_ids	Comma-separated list of cluster indices to include. Note the comma delimiter, which differs from the || used elsewhere.
cluster_distance_min / cluster_distance_max	Lower and upper bounds on centroid distance, used to focus on central or peripheral cluster members.
cluster_zoom	1 indicates that a single cluster is currently zoomed in the UMAP view. Included automatically when sharing a URL from a zoomed cluster so the recipient lands on the same view.

Metadata and Annotation Filters

Parameters that drive Metadata Search.

Parameter	Description
filter	||-separated list of annotation labels that clips must match.
filter_mode	any (default) or all. Combination rule for the values in filter.
labels_to_exclude	||-separated list of annotation labels that clips must not carry.
labels_to_exclude_mode	any (default) or all. Combination rule for labels_to_exclude.
label_types	||-separated list of label types to scope filter and labels_to_exclude to.
without_ann	true restricts to clips with no annotations of the active project, useful for prioritising labelling work.
search_clipid	Exact clip identifier. Surfaces a single clip by ID.
search_country	Filter clips by the recorded country using official two-letter country codes only. Format: ISO 3166-1 alpha-2 (e.g., Switzerland = CH, USA = US, Germany = DE).
left_hand_driving	true restricts to left-hand-drive recordings.
numeric_filter	||-separated list of numeric metric filters. Each filter has the form <metric_key>,<min>,<max>,<ordering>. Setting ordering to asc or desc makes the metric the active ranked mode.

Display Toggles

These parameters do not change which clips are returned, only how they are displayed in the result grid.

Parameter	Description
times	true shows time-range badges on each clip card when annotations are time-scoped.
with_metrics	true overlays per-clip model evaluation metrics on each card.
with_bev	true adds the bird’s-eye-view visualisation to each card.
n	Number of clips per result page. Defaults to 6 and is included in the URL only when changed from the default.

Worked Examples

The examples below show how a natural-language description of a slice maps to the corresponding URL.

Caption keyword search restricted to a dataset

“Find ambulance footage in MADS-1M.”

http://your-wheel-server/#&page=0&search=ambulance&data_source=MADS-1M&project_source=Alpamayo

Caption keyword combined with semantic retrieval and query rewriting

“Show me clips mentioning officer while semantically related to a construction zone, with rewriting on, in AV V1.”

http://your-wheel-server/#&page=0&search=officer&semantic_search_text=construction%20zone&semantic_extra_queries=road%20work%7C%7Cconstruction%20cones%7C%7Corange%20barrels&data_source=AV%20V1%20train%7C%7CAV%20V1%20validation&project_source=Alpamayo

Classifier filter with a probability threshold

“All clips that the Barrier gate classifier scores above 0.95.”

http://your-wheel-server/#&page=0&classifier_run_id=<classifier_run_id>&probability_expression=p%20%3E%200.95&project_source=Alpamayo

Replace <classifier_run_id> with the id of a trained classifier run (set automatically when you select a classifier in the UI).

Common Patterns

Working with URLs directly is most useful for:

• Sharing and bookmarking: paste the current URL into a tracker, a message, or a doc to give a colleague the exact slice you are looking at.

• Programmatic links: build a URL from a notebook or script when presenting a slice for review, for example after running a custom query through WheelClient and wanting to surface the same results in the browser.

• Adjusting filters by hand: tweaking a single parameter in the URL bar is sometimes faster than navigating the UI, especially when iterating on a long composed query.

For most interactive use, configuring the search through the UI panels is the natural path; the URL is the serialisation, not the primary interface.