> For the complete documentation index, see [llms.txt](https://docs.unitlab.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.unitlab.ai/documentation/data/embeddings-similarity-and-outliers.md).

# Embeddings, similarity, and outliers

Embedding tools help experts find the data worth inspecting. They do not replace ontology decisions, review, or source-aware judgment.

### Before you make the change

* Choose a source scope and compatible embedding space.
* Know whether the question is coverage, similarity, duplication, or anomaly discovery.
* Keep provenance and business metadata visible during interpretation.

![Find Similar result cohort](https://292810646-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FGjVLUz4wthGkGlRKM6rM%2Fuploads%2FmoLNHYRw2eyGq3iAMLTT%2Ffind-similar-results.png?alt=media\&token=dda7d9ce-1786-4420-96ba-4daca8bb2f6b)

*Nearest-neighbor results become useful only after a human checks source context, false positives, and the intended action on the cohort.*

### Understand the product behavior

The embedding view projects embeddable assets into a visual space and supports drag or crop-style region selection. Users can select a cluster and review its associated thumbnails before adding, excluding, or organizing the selected assets.

Useful curation patterns include:

* locating dense duplicate or near-duplicate regions;
* sampling from visually distinct clusters;
* finding outliers and rare conditions;
* comparing source domains;
* building a more diverse pilot dataset;
* selecting negative examples near a target class.

The two-dimensional plot is a navigation aid, not a guarantee that every nearby point is semantically identical. Operators should inspect the actual assets before turning a region into a dataset.

### Custom embedding spaces and vector search

The SDK can create named embedding spaces with a specified dimensionality and optional model name, upload vectors for an asset or a video frame, bulk-upsert vectors, search by a query vector, and optionally scope the search to a project or level.

This allows teams to use Unitlab’s curation interface while retaining control over the embedding model and vector space used for similarity operations.

### Search, duplicates, outliers, and UMAP

Unitlab includes visual similarity search, natural-language CLIP-style search, duplicate detection, outlier detection, quality-inconsistency checks, and UMAP visualization. These embedding and visual exploration experiences are available for image and video data.

### Turn vector evidence into a controlled cohort

{% stepper %}
{% step %}

#### 1. Inspect global distribution

Open Embedding view before zooming into a cluster so the local region has context.
{% endstep %}

{% step %}

#### 2. Choose a seed or region

Select a representative item, freehand region, or query vector.
{% endstep %}

{% step %}

#### 3. Review candidates

Compare neighbors, duplicates, or outliers against metadata and the native asset viewer.
{% endstep %}

{% step %}

#### 4. Separate signal from decision

Remove false positives and document the criterion used by the human reviewer.
{% endstep %}

{% step %}

#### 5. Version the accepted cohort

Tag, group, or publish the verified membership as required by the operating flow.
{% endstep %}
{% endstepper %}

### Decisions that affect production

| Decision           | Production guidance                                                                        |
| ------------------ | ------------------------------------------------------------------------------------------ |
| Embedding model    | Record the selected space because results can change across models or versions.            |
| Distance threshold | Use as a retrieval control, not as an unreviewed acceptance rule.                          |
| Duplicates         | Decide whether to remove, group, down-weight, or retain them for a documented reason.      |
| Outliers           | Review for rare value, corruption, distribution shift, or ingestion failure before acting. |

### Continue the operating flow

* Create a balanced dataset version.
* Use the cohort for annotation, review sampling, or model evaluation.
* Retain the embedding and source version in the decision record.