> ## Documentation Index
> Fetch the complete documentation index at: https://docs.dataerai.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Embedding map

> See a collection or a set of search results as a 2D similarity map of their images — visually alike pictures sit near each other, so clusters and outliers jump out at a glance.

The embedding map turns a set of images into a **2D similarity scatter**: Dataerai
places each image so that visually alike pictures sit close together and different
ones sit far apart. Clusters, duplicates, and outliers that are invisible in a
file list become obvious at a glance.

It runs on the image embeddings Dataerai already computes for content search, so
there's nothing to set up — open the map and Dataerai builds it for you.

<Note>
  The map plots only images you have **Read** access to the content of, and only
  assets that already have an image embedding. Assets you can list but not open,
  and assets with no picture, never appear on the map.
</Note>

## Map a collection

Every collection has a **Map** tab alongside **Data** and **Readme**.

<Steps>
  <Step title="Open the Map tab">
    Open a collection and select the **Map** tab. The first time, select
    **Generate map** — Dataerai computes the projection in the background and
    shows it when it's ready.
  </Step>

  <Step title="Explore the scatter">
    Each point is one image's thumbnail. **Drag** to pan, **scroll** to zoom, and
    **click** a thumbnail to select that asset (its detail opens in the sidebar
    when the asset is in the view you're browsing).
  </Step>

  <Step title="Refresh after changes">
    Select **Refresh** to recompute the map after the collection's images change,
    so the layout reflects what's in the collection now.
  </Step>
</Steps>

## Map your search results

On a search, switch the results from **List** to **Map** to see the matching
assets as a similarity scatter instead of a list. A picker lets you choose what
the map plots:

| Source       | What it maps                                                                                                                                                                                                                                                                             |
| ------------ | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Results**  | Exactly the assets your search returned — the precise, keyword/filter result set, laid out by visual similarity.                                                                                                                                                                         |
| **Semantic** | A visual (image-similarity) search of your query text. Use it when a keyword search misses pictures whose filenames don't mention the term — searching `flower` finds the flower images even if none are named "flower". The **Semantic** source appears only when you've typed a query. |

The **Map** view on the [advanced search](/discover/search) results works the same
way: toggle between **Graph** and **Map** to flip the matched assets between the
relationship graph and the similarity scatter.

## Choose a projection

The method picker controls how the high-dimensional similarity is flattened to
two dimensions. All three are deterministic — the same images always lay out the
same way.

| Method    | Best for                                                                                   |
| --------- | ------------------------------------------------------------------------------------------ |
| **PCA**   | A fast, stable overview that preserves the broad spread of the set. The default.           |
| **t-SNE** | Tight, well-separated local clusters — good for spotting near-duplicates and small groups. |
| **UMAP**  | A balance of local clusters and overall structure, useful for larger, more varied sets.    |

Switching methods recomputes the map for that view; the result is cached, so
returning to a method you've already run is instant.

## What you'll see while it builds

A map computes asynchronously. Depending on the state you'll see:

* **Computing the projection…** — the map is being built; it appears when ready.
* **No images to map** — none of the assets in scope have an image embedding yet.
* **Nothing to map** — the search returned no results to plot.
* A map that's **taking longer than expected** — a one-click **Keep waiting**
  lets you keep polling if a large set is still computing.

<Note>
  Each map is yours alone — it's computed against your own access, so two people
  opening the same collection's map each see only the images they can open. There's
  a small per-person limit on how many maps can compute at once, so a map may wait
  briefly if you've just kicked off several.
</Note>