Skip to content

Gradients ​

Gradients treat a space as a continuous field rather than a set of clusters. The backend fits a set of feature axes over the space, and Mantis renders those axes as color maps, field lines, hotspots and 3D terrain, or maps them onto point color and point size. Six panels make up the family: the Gradients panel and three map views, two of which have a companion controls panel.

Where to find it ​

View > Exploration > Gradients opens the Gradients panel. From there, three buttons open the map views, and two of them also open their controls panel:

  • Open Cross-Axis View
  • Open Arithmetic View, plus Arithmetic Controls
  • Open Feature Scatterplot, plus Scatterplot Controls

Each view opens in the tabset that holds the Map tab, and only one copy of each can exist. Clicking a button again focuses the tab you already have. All three views embed Map (Legacy), because the gradient geometry is rendered inside that scene.

Features have to exist first ​

Everything here reads a list of features computed for the space. If they have not been computed yet, the panel says so and offers a retry. The Recompute button at the bottom of the panel triggers the job and lets you pick the method:

  • SVAE, the default and the better result. Mantis polls for it for up to about 25 minutes.
  • PCA, faster, polled for about 5 minutes.

The Gradients panel ​

The panel is titled Gradients Explorer. Feature Selection picks the active feature and toggles the overlay on or off. Once it is active, four more sections appear:

  • Field Line Controls: show field lines, animate them, and set arrow density. Field lines are drawn as wavy lines with arrowheads that travel along them.
  • Concentration Areas (Hotspots): show hotspots, set how many, set the density and feature thresholds, and choose whether hotspots are found by density, by feature value, or both. Two more toggles compact the cluster and feature labels drawn inside them.
  • Visualization Mode: 2D or 3D. In 3D the gradient becomes terrain, with height exaggeration, surface opacity, wireframe and contour-line controls.
  • AI Insights: generates a written summary of the active feature from a sample of the space's point text.

Below those sit three more tools:

  • Semantic Grouping switches the current map between Spatial and Semantic layout. In semantic mode, points are regrouped by their dominant feature axis, and a slider moves each group between collapsed, original and expanded spacing. This changes the map in place.
  • Path Analysis shows the result of the last path you drew across the map while gradients were active, including how feature values change along it.
  • Advanced Options enables per-feature toggles, a search radius around the current selection, and an analysis of the features in that radius that refreshes as you move the selection or change the radius.

Cross-Axis ​

Cross-Axis compares two features by recoloring the points, with no surface and no plane. Pick Feature 1 (red) and Feature 2 (blue) from the two lists in the control bar, then click Compute. Each point is colored by blending its normalized values for the two features: red where the first dominates, blue where the second does, purple where both are high, and muted gray where neither is. The legend sits at the right of the bar.

Once a comparison is on screen, two more buttons appear: Export, which downloads a PDF brief of the comparison, and Clear, which removes the coloring.

Arithmetic ​

Arithmetic builds a new field out of the ones you have. The map tab itself is just a status strip and the map; all controls are in Arithmetic Controls.

  1. Build the expression. Each row is a weight times a feature. Use + term to add rows and the trash icon to remove them. Weights accept negatives, so you can subtract one feature from another.
  2. Click Compute. The derived field renders on the Arithmetic map with a diverging color scale: red where the combined value is high, blue where it is low, near-white where it is close to zero. The applied expression and its terms are listed underneath.
  3. Switch to 3D to see the field as terrain. The 3D button stays disabled until a computation has produced a height matrix.
  4. Adjust the surface with height exaggeration and surface opacity, and turn on the wireframe overlay. These take effect in 3D.
  5. Export downloads a PDF brief of the derived field, and Clear removes the overlay.

Feature Scatterplot ​

Feature Scatterplot keeps every point where the embedding puts it and encodes features into the marks instead. In Scatterplot Controls:

  • Size by picks the feature that drives point size. Points below the 5th percentile shrink to 40 percent of the base size; points above the 95th grow to 160 percent.
  • Color by picks the feature that drives point color, on a continuous low-to-high scale shown as a legend bar. Leave it empty to keep cluster colors.
  • Reset clears both.

Either control can be used alone.

How the overlays are scoped ​

Each view owns its own overlay slot, so a Cross-Axis coloring never leaks into the Arithmetic tab. Closing a view clears its overlay. This is also why only one tab of each view may exist at a time.

Limits ​

  • The gradient plane, field lines, hotspots and 3D surface render only in the legacy scene, which is what the three views embed. On the default Map an active gradient does one thing: in 2.5D it lifts points to the gradient's height. It draws no plane, no lines and no hotspots.
  • Recompute is a long backend job. The panel polls and reports progress; it does not block you from working elsewhere in the meantime.
  • Cross-Axis needs two different features. The Compute button stays disabled until both are chosen and they differ.
Read nextMap (Legacy)The renderer every gradient view is built on.

Gradient Compass actual walkthrough ​

Open Gradient Compass inside Gradients Explorer to move from an actual source toward a validated contextual direction. Choose a stored source-feature floor and minimum domain and hop similarities; compare the constrained route with the same search without those constraints. Inspect the actual source scores and review the comparison before retaining a route. Reloaded historical settings need a fresh comparison against current sources.

Actual source-space comparison, reviewed route retention and reload in the native Gradient Compass.

This edited recording holds genuine application screenshots for readability. It uses actual production Compass components, contextual training, authenticated Django requests, CSRF enforcement and PostgreSQL. A validated criterion is trained from invented original examples with independent validation examples. The constrained route preserves a stored feature floor of 0.7 and stops short of the 0.8 target; the unconstrained route reaches the target while violating that floor. A partial bounded search does not prove no feasible route exists. Semantic feature scores do not certify scientific evidence quality.

Actual tradeoff: constrained partial route versus an unconstrained target-reaching route with a feature-floor violation.

The verification receipt records source inspection, reviewed retention, reload and reuse, actual infeasible-start 422, viewer-write and missing-CSRF 403, stale source-score/axis-provenance 409, and exact retry of an already committed save after that change. Restoring the owned fixture's score restores comparison. Original rows, text, vectors, map geometry and configuration retain matching before/after hashes. The native splitter is resized to 357 pixels in light and dark themes without horizontal overflow.

The verified API source is main 7ff25f619b983f8bba3accef94da6a647c28f49f; frontend runtime source bytes match main ddc85b8f6f8c5a47c62ed04583dc11e4cebd185a. The fixture contains eight invented sources/vectors and an explicitly input-derived deterministic intent-embedding provider. Ranking and route results are actual services. Seeded sign-in, the adjacent map-selection/store bridge and enclosing Gradients header replace the full deployed Next.js/OAuth/map shell; production GradientCompassTool, GradientCompass, FlexLayout and Mantis styles are used. The owned database uses synchronized current models; this is not historical migration proof or a customer-outcome claim. A source feature score was deliberately changed and restored for the freshness check; its unrelated notification enqueue uses an owned in-memory broker.