AI powered semantic search

Brandkit now supports semantic search (Wikpedia definition) as a configurable option. Currently in BETA. It must be enabled in your Account by the Brandkit team - so contact us to enable it - for you to try out.

(Fig: Example of a semantic search for “Romantic Dog” using a vague’ish sentence/query)

Essentially once your Assets are classified, tagged, and described, we run AI again to embed your content metadata in a vector database, which is then used to power semantic (aka natural language) searches.

Huh? Plain english please!.

What this means for users is that they can search using vague’ish phrases or sentences and get sensible search results.

How it works

When you upload or create a new Asset (if Automation is enabled), a description and multiple tags will be auto-generated for you (provided you meet the Requirements below).

Next Brandkit will automatically generate what is called Embedding Content. This Embedding Content is a combination of asset name, description, tags and other metadata saved as a single string of text.

(Fig: Screenshot showing Embedding Content in the edit metadata page for this photo)

This “Embedding” text is optimised for semantic search and is saved to a special kind of database (think “search index”) called a vector database, where each string of text is represented by score in 3d space (conceptually).

When entering a semantic search query, the system will convert that search query into a score and then look for Assets with the closest score in the vector database.

The search results will then be displayed with the nearest match first, followed by other assets in order of relevance (based on each asset’s vector score).

(Fig: Screenshot showing the nearest match to the semantic search query first)

So in the screenshot above we can see that the dog with the Rose between his teeth is the first Asset found - or in other words , has a score closest to the score of the search query “I need picture of a romantic dog for valentines day”

Note - that in semantic search results there are no other sorting options - because semantix results are always sorted by relevance (i.e. score) to the search query (also a score).

Requirements

For Semantic Search to work effectively your Assets must be properly tagged and described.

To ensure that this is done first enable Auto Tagging and Auto Descriptions in your account’s Automation settings.

1. Navigate to Admin > Settings > Automation
2. Enable automatically tag with file metadata on ingestion
3. Enable Automatically tag with AI on Ingestion
4. Enable Automatically describe with AI on ingestion
5. Ensure you have sensible (and tested) prompts setup for both tagging and descriptions.

Note: that these settings will only apply to new Assets on creation. For any existing Assets you should batch run Auto-Tag, Auto-Describe across all Assets, and once that is completed rune Batch Update Embeddings across all Assets (this will populate the vector database used for semantic search).

Note also: it is important that these steps are followed in order. DO NOT run embedding if the Assets do not have completed tags and descriptions. And if you add new descriptions, tags (or other metadata) to an asset you should update the embeddings for the Asset.

Please contact us if you have any questions or to us feedback on this new feature.

Happy branding :)