Vector search and Embeddings saving

[somasekharkakarla]

Problem

Hi @takameyer and @kneth

Currently, LLMs are in high demand. SLMs are also gaining their strength.
I want to request native support for feature vector search in Realm. This would enable developers to store embeddings (e.g., image, text, or audio vectors) and perform similarity searches directly within Realm, without relying on external vector databases.

Proposed Capabilities:

Support for storing float[] or double[] fields as feature vectors.

Native querying by similarity metrics such as:

Cosine similarity

Euclidean distance

Ability to combine vector queries with existing Realm filters (e.g., by category, user, or timestamp).

Efficient indexing to allow fast K-nearest-neighbors (KNN) queries on mobile.

Compatibility with React Native, iOS, Android, and Node.js SDKs.

Use Cases:

Mobile AI applications: On-device image, audio, or text search using embeddings.

Recommendation systems: Quickly find similar items or content for users.

Hybrid workflows: Store metadata in Realm while performing fast vector queries without a separate vector DB.

Benefits:

Reduces complexity by eliminating the need for external vector databases.

Speeds up development of AI-powered applications on mobile and web.

Expands Realm’s capabilities for next-generation apps using embeddings.

Offline first and privacy.

Priority: High — Feature vector search is increasingly critical for AI/ML applications that store and query embeddings on mobile and cloud.

Solution

No response

Alternatives

No response

How important is this improvement for you?

Dealbreaker

Feature would mainly be used with

Local Database only

[kraenhansen]

Thanks for your interest in improving the library 💙

You might not have noticed this, but the library is deprecated and as such we won't be expanding it with new features. Neither @takameyer nor @kneth are working on this project anymore, so I wouldn't expect an answer from tagging them.

With those formalities out of the way, I'll surface this request internally. I do remember that we did experiments with a vector search index on numeric arrays at some point, I'll try to dig up a link for you.

[kraenhansen]

For anyone interested in this, here's a branch experimenting with adding this KNN to Realm Core: as/vector-search.