Projection
The Projection panel answers one question: where on this map would this piece of text land? It trains a small neural model on the current map, then maps any text you type to an (x, y) coordinate in that map’s space and lets you drop a pin there.
It is not a dimensionality-reduction workbench. Choosing UMAP, t-SNE or PCA and tuning their parameters happens in Space Creation > Refine > Dimensionality Reduction Phase, or at space-creation time.
Where to find it
View > Machine Learning > Projection. The panel is headed “Text Projection Search”.
It works on the map you currently have open; the model is per map, not per space.
Model status
A colored dot and label at the top tell you where the model stands:
| Dot | Label | Meaning |
|---|---|---|
| Gray | Checking model… | Status request still in flight |
| Green | Projection model ready | A trained model exists and its dependencies are installed |
| Red | Projection model not available | No model, or the backend is missing its dependencies |
Refresh re-checks the status. When no model exists, a Train Model button appears next to it.
Below the pill, a Model Status Details block reports whether dependencies and a trained model are present, whether training completed, a percentage, a status message, and any error the backend returned.
Training a model
- Click Train Model. Training runs on the backend for this map, with a maximum of 100 epochs and hidden layers of 128, 64 and 32 units. You cannot change those from the panel.
- Watch the log. The panel opens a Training Logs view fed by a live log socket, timestamped and colored by level. Toggle it with Show Logs / Hide Logs, and empty it with Clear Logs.
- Wait. The panel polls the status endpoint every 3 seconds. Training a map takes minutes, not seconds.
- Done. On success the status pill turns green and the text box unlocks.
Heads up
Polling stops after 10 minutes and the panel reports “Training timeout - please check status manually”. That is the panel giving up on watching, not the backend giving up on training. Click Refresh later to see whether the model finished.
Training failures surface as a red banner with the backend’s error message, and both the start and the outcome are written to your User Logs.
Projecting text
- Type into the text box. It stays disabled until a model is ready.
- Click Project to Coordinates, or press Enter.
- Read the result. The Projection Result block shows the raw coordinates the model returned and the scaled coordinates after the map’s scale factor is applied, both to four decimal places, along with the input text.
- Click Drop Pin Here to place the pin. The text box clears and the result block closes.
Dropped pins
Pins render on the map as red markers. In 2.5D they are drawn as signposts, with a pole and a base dot; in 2D they sit flat at the coordinate.
The Dropped Pins list under the result shows each pin’s text, its raw and scaled coordinates, and the time it was dropped. Each pin has an X to remove it, and Clear All empties the list for the current map.
Pins are held per map in browser memory. They are not saved to the space and do not survive a reload, and other people looking at the same space do not see them.
Limits
- One model per map, with fixed hyperparameters. Retraining replaces it.
- One text input at a time; there is no batch projection.
- The model returns a single 2D coordinate. There is no confidence score and no nearest-point lookup.
- The panel does not modify the map. Nothing is re-embedded or re-reduced.