Space Creation

Space Creation is the panel for changing a space after it exists. You can re-run any single phase of the synthesis pipeline over the current map, re-embed it on a different set of fields, or add a whole new map to the space from data it already holds.

Everything it does is a backend job. Nothing here happens instantly, and nothing here uploads new data.

Where to find it

View > Machine Learning > Space Creation.

The panel has a tab strip across the top: Refine, Add Map, Embed Field, Composite Map, Upload JSON, Statistics. It opens on Refine.

Heads up

Upload JSON and Statistics currently render nothing. The tabs are listed but the panel body has no case for either, so selecting one leaves the content area blank.

Refine

Pick one phase, configure it, and invoke it. The section header reads “Select a phase to invoke”, and each phase is a radio button that reveals its own parameter block.

Phase	What it does	Parameters
Re-Embed	Recomputes embeddings for the map	Select Semantic Columns: which metadata columns feed the embedding
Dimensionality Reduction Phase	Recomputes 2D coordinates from the embeddings	Reduction method plus its parameters
Clustering Phase	Rebuilds the cluster structure	Clustering method plus linkage, distance metric and toggles
Cluster Labelling Phase	Regenerates cluster labels from cluster content	None
Point Re-titling Phase	Regenerates point titles from metadata	Which columns to use, plus an Include Title Field checkbox

Click Invoke Phase to start. The button reads “Processing…” while the request is in flight.

Reduction parameters

The Reduction Method radio group offers three methods, each with its own fields:

• UMAP: Number of Neighbors (minimum 2) and Minimum Distance (0 up to but not including 1).
• t-SNE: Perplexity (5 to 50) and Max Iterations (minimum 250).
• PCA: Number of Components (2 or 3).

Reduction targets 2D. The map’s 2.5D view is a rendering mode, not a third reduced dimension. There is no distance-metric control here, even though the backend accepts one for UMAP.

The create-space forms offer a wider menu than this panel does, including PCA+t-SNE and PCA+UMAP, so if you need one of those, choose it when the space is created.

Clustering parameters

Two methods: Hierarchical Clustering and HDBSCAN.

Hierarchical exposes a linkage method (Ward, Complete, Average, Single, Weighted, Centroid, Median), a distance metric (Euclidean, Manhattan, Cosine, Correlation, Hamming), and two checkboxes: Automatic Optimal Depth Detection, and Ignore References, which excludes citation-style reference points from determining the cluster structure.

HDBSCAN exposes a Minimum Cluster Size slider from 2 to 50 and the same distance metric choice.

Watching a phase run

Refinement phases are asynchronous backend jobs. The panel does not block, and the map does not update at the moment you click.

After a successful invoke, a Show Logs / Hide Logs toggle appears at the bottom of the panel over the map’s synthesis log. Use it to follow the job; the map picks up the result when the job finishes.

1. Select the Refine tab and pick the phase you want.
2. Configure its parameters. The column pickers load the map’s metadata columns on demand.
3. Click Invoke Phase.
4. Click Show Logs and watch the job to completion.

Note

If the current map was imported from an H5AD file, Refine refuses to run and shows “Imported H5AD source map” instead. Coordinates, cluster assignments, labels and colors come from the AnnData file and are preserved. Create a derived representation map first if you want to re-embed, reduce, recluster or relabel.

Add Map

Adds a second map to the space you are already in. It does not upload a file, and the create-a-space flow does not apply.

The tab opens with an Alternative View switch. Left off, it shows an “Aggregation View” placeholder with no controls. Turn it on for the working flow:

1. Click Load Metadata Columns. This fetches the current map’s numeric metadata columns.
2. Narrow the list. Type in the search box to filter, or click the ABC button to flip it to .* regex mode and filter by pattern. The counter reads “Showing N of M columns”. With a search active you also get Select All Filtered and Deselect All Filtered.
3. Tick at least one column. The submit button stays disabled until you do.
4. Click Submit Selected Columns (N). Mantis builds a new map from those columns and redirects you to it.

Embed Field

Re-embeds a map on a chosen set of semantic fields, with per-field weights. Headed “Re-Embed with New Fields”.

Choose a Map from the dropdown of maps in the space, tick the semantic fields to include, set each field’s weight, and choose an embedding mode: Concatenate text joins the fields into one string before embedding, Separate embeddings embeds them separately. Click Re-Embed to start.

Fields come from the map’s semantic columns as already loaded into the workspace; if a map has none, the panel says “No semantic fields found for this map.” On success it reports “Re-embedding started. Check synthesis progress for updates.” This overwrites the selected map’s embeddings.

Composite Map

Builds a new map from an existing one’s semantic fields instead of overwriting it. Headed “New Composite Map”.

Pick a Source map, choose fields and weights the same way as Embed Field, give the new map a Map name, and optionally tick Skip clustering (faster, no cluster coloring). Create Composite Map returns the new map’s id, and the map appears once synthesis completes.

Limits

• Upload JSON and Statistics are dead tabs in the current build.
• Nothing here is undoable. Re-running a phase overwrites the map’s embeddings, coordinates, clusters, labels or titles in place.
• The panel gives no completion notification. Use its log toggle, or the Synthesis Progress panel.

Read nextProjectionTrain a text-to-coordinate model on a map you refined.