Skip to content

Summarizer ​

The Summarizer, titled Inspector+ / Summarizer in the app, watches your selection and analyzes it thirteen different ways. Select two or more points and it generates a written summary, charts the fields they share, ranks their keywords, finds outliers and correlations, and can save the result as a vignette. Nothing here is a chat: the panel reacts to the selection on its own.

Where to find it ​

Open View > Inspection > Summarizer, or select two or more points on the map and use the Summarize Selection point action. That action needs a selection of at least two points, and it only appears in spaces that have semantic (text) columns.

The panel needs point metadata to have finished loading.

How it fills ​

Everything is driven by the current selection. With nothing selected the panel reads "Select points or a cluster to see pre-analyzed insights."

The written summary is generated from the selected point ids about 1.5 seconds after the selection settles, so quick lasso adjustments do not each trigger a request. Generated summaries are cached, so revisiting a selection you already summarized is instant.

Every generated summary is also saved as a snapshot. The Saved (N) button in the header lists them with their point count and timestamp; picking one shows that analysis again, which is useful after you have cleared the selection.

The tabs ​

Dashboard is the default. For a multi-point selection it stacks header counts (points, clusters, properties), the AI summary, bar charts of the categorical fields, min/mean/max tables for the numeric fields, cluster distribution, a per-cluster radar, an image grid and a keyword cloud. Hovering a chart segment highlights the matching points on the map. Two buttons sit under the summary: Continue in Composer, which opens the Composer with this analysis attached, and Save as Vignette, which writes the summary to a new vignette named after its first words and the timestamp. A single selected point gets a compact detail view instead.

Summary is the written analysis on its own. Cluster and Tree selections show the stored cluster summary; a selection of two or more points shows the generated summary alongside enriched keywords, the date range, per-field numeric statistics and representative points; a single point shows its details, not a generated summary.

Table lists the selected points as rows, 25 per page, with title, cluster and the metadata keys found on the first ten points. Click a header to sort, hover a row to highlight the point on the map. Embeddings, connections, links, layer and a few other internal fields are excluded.

Images pulls every metadata value that resolves to an image and shows them as a grid, labelled with their point and field name. Click one to expand it.

Distrib. charts one field at a time: a histogram for numeric fields with a Bins control, a Top N bar chart for categorical ones. Hovering a bar highlights those points on the map; clicking it narrows the selection to them.

Compare groups your selection by a categorical field, then compares group averages across up to six numeric fields as grouped bars and a radar. It needs at least one categorical and one numerical field, and tells you which is missing. Averages use finite observations only: a group with no finite values for a field gets a — in the table and no bar, not a zero. Its radar profile is replaced with an explanation naming the missing fields. Genuine measured zeros remain visible, and group counts include every point in the category.

Contrast compares two bags rather than the selection. Pick a bag for each side, and Mantis computes cluster and keyword overlap in the browser, asks the backend for group characteristics and feature relevance (Cohen's d, p-values, Cramér's V, ranked importances), then generates a three-paragraph narrative: a profile of each group and the differences between them. Both bags need at least two points.

Trends plots one numeric or date field against another for the selected points, with selectors for each axis. Dates are the default x axis when the space has them.

Radar averages up to six numeric fields per cluster within the selection and overlays the clusters on one normalized radar chart. It needs at least three numeric fields. Radar and the Dashboard radar use finite observations only. A cluster missing every finite observation for any displayed field gets a named explanation instead of a fabricated zero or polygon. Measured zeros remain valid; different points can supply different fields in a complete cluster profile. Existing cluster ranking and display limits are unchanged.

Insights scans the selection for five kinds of pattern and issues a card for each one it finds: a dominant category, a skewed distribution, high variance, outliers and strong correlations. Hovering a card highlights the points behind it; outlier and dominant-category cards add a button that narrows the selection to those points. This tab runs entirely in the browser.

Education generates study material from the selection. Choose Flashcards or MCQ Quiz, set the item count (3 to 20) and the difficulty (easy, medium, hard), then press generate. Flashcards flip and you grade yourself; quiz questions are scored, and after the results screen you can Restart Session, start a New Session, or Review Incorrect to redo just the ones you missed. Decks can be saved to the chat session, and there is a fullscreen mode.

Gradients narrates the last path you drew with the Path tool: it runs path analysis over all ten gradient axes and returns an AI narrative of the semantic journey, a sparkline and delta per axis, and the ideas found near the path. It needs a path of at least two points, and works whether or not gradients are toggled on in the map.

Concepts builds a small concept graph from the selection: a center node for the selection, phrase and concept nodes around it, sized by how many points match. Click a node to focus it and highlight its points, drill down and use Back to walk out again, Create Bag (N) to keep the highlighted points, and the wayfinding selectors to compute a route between two concepts. It can also narrate itself with an interactive transcript.

Concept overlays and routes belong to the map supplying this panel's selection; using Merge or closing the panel does not reset another map's overlays. Route anchors support dragging, or keyboard routing: focus an anchor and press Enter or Space, then activate the destination anchor. Escape cancels the first anchor. In the session graph, activate two connection-preview buttons to inspect a missing connection. Playback uses stable graph positions; dragging nodes and saved snapshots still override those positions. To preview a route detour without dragging, activate a gap, then choose a candidate concept.

The Study view grounds tutor feedback in your selected sources. If those are insufficient, live research uses the backend's minimum six search iterations (which can take longer than a local hint). Research does not attach scraped text to the current space or replace its Input Text. If research is unavailable, the tutor explicitly labels its response as unverified background guidance.

Narrate and Narrate Tour use the same browser speech engine as the Mind Map panel: starting one replaces the other. A pending start shows Starting narration... and can be stopped. Playback failures, including a start that takes more than ten seconds, appear in the panel.

  1. Select a set of points, or a cluster, or a bag.
  2. Wait a moment on Dashboard. The AI summary appears after the selection settles.
  3. Move to the tab that answers your question: Distrib. for the shape of a field, Insights for anomalies, Compare for group differences.
  4. Push the result somewhere. Save as Vignette keeps the write-up with the space; Continue in Composer carries it into a conversation.

Limits ​

  • A generated summary needs at least two points. One point gets a detail view; zero points get the placeholder.
  • There is no length, tone or format control for the generated summary. The length slider lives in the Reader, which is a different panel.
  • Several tabs need particular field types and say so plainly when they are missing: Compare needs categorical and numeric fields, Radar needs three numeric fields, Trends needs numeric or date fields, Contrast needs two bags of at least two points each.
  • Snapshots live in the browser session. They are not saved with the space; save a vignette for anything you want to keep.
Read nextVignettesWhere Save as Vignette puts the write-up, and how to share it.