Databases have learned a small set of default views. Table, board, calendar, gallery, timeline, list. These are now so familiar that we barely notice their assumptions. A table says fields matter. A board says status matters. A calendar says time matters. A gallery says visual identity matters. A graph says explicit links matter.

There is one view missing from this set: a view where items arrange themselves by what they are like. Not by a hand-written tag, not by a single status column, not by a folder. By resemblance.

Call it Similarity View. It is a map for a collection. Nearby items are related. Dense areas are themes. Lonely items are outliers. The point is not to make embeddings visible for their own sake. The point is to give ordinary collections a new sense: spatial perception of meaning.

The missing view

A task is rarely just todo, doing, or done. It may be almost implemented but poorly described. It may be blocked but strategically important. It may be urgent but vague. It may be marked complete while sitting semantically next to unfinished work. It may be one of six near-duplicates spread across three projects.

Current views hide this because they force work into one dominant axis. A kanban board is useful when workflow stage is the main question. But many project questions are not workflow questions. They are shape questions: what is clustering, what is isolated, what is repeated, what is weird, what belongs together, what does not fit.

Similarity View is a way to ask those questions without first knowing the right tag, query, or schema. It is less like sorting a spreadsheet and more like stepping back from a wall of sticky notes until patterns become visible.

What it is

Similarity View is a collection view where each item becomes a point or card on a two-dimensional map. The layout is generated from embeddings and metadata. Text, comments, labels, owners, status, age, links, and activity can all contribute to where an item appears. The user does not need to know this. They only need the simple rule: things near each other are meaningfully related.

The view can be used for tasks, notes, bugs, support tickets, product feedback, research snippets, candidates, leads, documents, bookmarks, and design references. It is not a replacement for every table or board. It is a complementary view for moments when the collection is too large, too fuzzy, or too emergent for rows and columns to explain.

Why now

The idea has existed in research for decades in different forms. What changes now is availability. Embeddings are cheap enough, good enough, and general enough to become an everyday interface material. They no longer need to live only inside search, recommendation, or ML debugging tools.

Also, collections have changed. The modern project is not a clean set of tasks. It is a pile of Slack threads, support messages, design notes, customer calls, specs, bugs, screenshots, docs, and AI-generated drafts. People need a way to see the shape of that pile before they know how to classify it.

That is the practical opportunity. Similarity View turns machine-readable resemblance into human-readable layout. It makes the latent structure of a collection available as a surface for browsing, selecting, naming, and acting.

The interface

The interface should feel mundane. A user opens a database and chooses View → Similarity. The collection becomes a map. Items can still be filtered, searched, colored, sized, opened, edited, and selected. The difference is that position now carries meaning.

• Position shows resemblance. Items close together are related by content, context, or chosen lens.
• Color can show status, owner, priority, sentiment, project, or source.
• Size can show urgency, impact, effort, recency, confidence, or number of comments.
• Density shows themes. A crowded region means many items share something.
• Distance shows surprise. A faraway item may be unique, misplaced, stale, or important.
• Lasso selection turns a cluster into an action: tag it, summarize it, assign it, make a project, merge duplicates, archive noise.

The killer interaction is probably not panning and zooming. It is lasso → name → act. Select a cluster of customer complaints and name it “onboarding confusion.” Select a knot of vague implementation cards and turn them into a planning session. Select the outliers and ask why they do not fit.

collection
  ↓ embed items
semantic map
  ↓ lasso a region
cluster becomes named work
  ↓ act
project / tag / owner / summary / archive / duplicate merge

What it reveals

The view is useful because it reveals things the user did not explicitly encode. A table can only show columns that already exist. A board can only show statuses someone already chose. A Similarity View can show structure before the team has agreed on a taxonomy.

• Clusters: emergent themes that deserve labels, owners, or projects.
• Duplicates: separate items that probably describe the same need, bug, complaint, or idea.
• Gaps: missing work between two dense areas, or a region that should exist but does not.
• Outliers: strange items that are isolated, mislabeled, obsolete, unusually broad, or unusually valuable.
• Bridges: items that connect two clusters and may represent hidden dependencies or synthesis opportunities.
• Status mismatches: items marked “done” that sit inside unfinished territory, or “todo” items that are actually near ready-to-ship work.

This is especially valuable for prioritization. Priority is usually treated as a field. But prioritization is relational. The question is rarely “what is the highest number?” It is “what changes if we act on this cluster, ignore that region, split this item, or close these duplicates?” Similarity View makes those tradeoffs visible.

Related design research

There is a serious design-research lineage behind this, even if the product should not feel academic. The closest ancestors are spatial hypertext, sensemaking, visual analytics, and embedding visualization.

• Spatial hypertext studied how people use position, grouping, and visual arrangement to make implicit relationships visible. Frank Shipman and Cathy Marshall’s work on spatial hypertext and systems like VIKI is directly relevant because it treats space as a medium for weak, emerging structure rather than only explicit links. See Spatial Hypertext: An Alternative to Navigational and Semantic Links.
• Personal collection workspaces like Visual Knowledge Builder explored how people collect, annotate, and organize digital materials visually. The paper Supporting personal collections across digital libraries in spatial hypertext is a useful precedent for collections as visual workspaces.
• Semantic interaction in visual analytics studies how users can directly manipulate spatial layouts while the system infers analytical intent and updates the underlying model. This is important if Similarity View becomes editable rather than static. See Semantic Interaction for Sensemaking.
• Sensemaking systems study how people move between foraging for information and organizing it into understanding. Recent systems like Sensecape combine LLMs with multilevel exploration, which is close to the “map, summarize, cluster, act” pattern.
• Embedding visualization tools already exist for ML and data analysis. Apple’s Embedding Atlas, WizMap,Embedding Comparator, and the Collection Space Navigator show how large embedding spaces can be explored with clustering, labels, cross-filtering, and projections.
• Latent-space creative tools like VideoMap show that latent spaces can support creative browsing and editing, not only search or classification.

The research gives confidence that the ingredients are real. But the product opportunity is simpler than the literature: make this a normal view option inside everyday collection tools.

What is new

The new part is not projecting embeddings onto a canvas. Researchers and ML tools already do that. The new part is treating this as a default collection primitive, next to table and board, for normal people managing normal messy work.

That means the design bar changes. It cannot look like an ML demo. It cannot require explaining UMAP, t-SNE, cosine distance, or vector databases. It has to behave like a view. It has to preserve the expectations users already have: filters work, fields work, selection works, permissions work, links work, undo works, comments work.

The translation is the important move: from embedding visualization as analysis tool to similarity layout as everyday UI. Once phrased that way, the idea becomes much easier to transmit.

A studio brief

If this were a studio project, I would prototype it around customer feedback first. Feedback has obvious clusters, painful duplicates, and high value. It also avoids the philosophical trap of “what is a task?” The user imports one thousand feedback items and immediately sees the terrain of complaints, requests, praise, and confusion.

• Start with a feedback collection: support tickets, app reviews, interview notes, sales call snippets.
• Generate a Similarity View with automatic cluster labels.
• Let the user lasso a cluster and create a product theme.
• Show representative items, edge cases, duplicates, and outliers inside that cluster.
• Let the user turn the cluster into a project, roadmap item, Jira epic, Linear project, or Notion page.

Then bring it back to project management. A product team could see that a cluster of “small bugs” is actually one onboarding problem. A manager could see that a supposedly active project is only a cloud of vague cards. An individual could see that half their todo list is the same unresolved decision wearing different names.

A careful name

“Embedding View” is accurate but too technical. “Map View” is friendly but too generic. “Atlas” is beautiful but maybe too grand. The cleanest category name is Similarity View.

It names the layout principle, the same way Calendar names time and Board names workflow. It is simple enough to sit in a view picker without explanation.

The compressed pitch:

Similarity View
A collection view where items arrange themselves by resemblance,
revealing clusters, duplicates, gaps, and outliers that tables and boards hide.

That is probably the whole idea. Not a new theory of work. A new sense organ for collections.