ObjectBox 4.0 is an on-device vector database allowing Android and Java developers to enhance their apps with local AI capabilities. A vector database facilitates advanced vector data processing and analysis, such as measuring semantic similarities across different document types like images, audio files, and texts. A classic use case would be to enhance a Large Language Model (LLM), or a small language model (like e.g. the Phi-3), with your domain expertise, your proprietary knowledge, and / or your private data. Combining the power of AI models with a specific knowledge base empowers high-quality, perfectly matching results a generic model simply cannot provide. This is called “retrieval-augmented generation” (RAG). Because ObjectBox works on-device, you can now do on-device RAG with data that never leaves the device and therefore stays 100% private. This is your chance to explore this technology on-device.
Vector Search (Similarity Search)
With this release it is possible to create a scalable vector index on floating point vector properties. It’s a very special index that uses an algorithm called HNSW. It’s scalable because it can find relevant data within millions of entries in a matter of milliseconds.
We pick up the example used in our vector search documentation. In short, we use cities with a location vector to perform proximity search. Here is the City entity and how to define a HNSW index on the location:
| 1 2 3 4 5 6 7 | @Entity data class City( @Id var id: Long = 0, var name: String? = null, @HnswIndex(dimensions = 2) var location: FloatArray? = null ) |
Vector objects are inserted as usual (the indexing is done automatically behind the scenes):
| 1 2 3 4 5 6 | val box = store.boxFor(City::class) box.put( City(name = "Barcelona", location = floatArrayOf(41.385063f, 2.173404f)), City(name = "Nairobi", location = floatArrayOf(-1.292066f, 36.821945f)), City(name = "Salzburg", location = floatArrayOf(47.809490f, 13.055010f)) ) |
To perform a nearest neighbor search, use the new nearestNeighbors(queryVector, maxResultCount) query condition and the new “find with scores” query methods (the score is the distance to the query vector). For example, let’s find the 2 closest cities to Madrid:
| 1 2 3 4 5 6 7 8 9 | val madrid = floatArrayOf(40.416775f, -3.703790f) val query = box .query(City_.location.nearestNeighbors(madrid, 2)) .build() val results = query.findWithScores() for (result in results) { println("City: ${result.get().id}, distance: ${result.score}") } |
Vector Embeddings
In the cities example above, the vectors were straight forward: they represent latitude and longitude. Maybe you already have vector data as part of your data. But often, you don’t. So where do you get the vector emebeddings of texts, images, video, audio files from?
For most AI applications, vectors are created by a embedding model. There are plenty of embedding models to choose from, but first you have to decide if it should run in the cloud or locally. Online embeddings are the easier way to get started and great for first testing; you can set up an account at your favorite AI provider and create embeddings online (only).
Depending on how much you care about privacy, you can also run embedding models locally and create your embeddings on your own device. There are a couple of choices for desktop / server hardware, e.g. check these on-device embedding models. For Android, MediaPipe is a good start as it has embedders for text and images.
Updated open source benchmarks 2024 (CRUD)
A new release is also a good occasion to update our open source benchmarks. The Android performance benchmark app provides many more options, but here are the key results:
CRUD is short for the basic operations a database does: create, read, update and delete. It’s an important metric for the general efficiency of a database.
Disclaimer 1: our focus is the “Object” performance (you may find a hint for that in our product name 🙂); so e.g. relational systems may perform a bit better when you directly work with raw columns and rows.
Disclaimer 2: ObjectBox delete performance was cut off at 800k per second to keep the Y axis within reasonable bounds. The actually measured value was 2.5M deletes per second.
Disclaimer 3: there cannot be enough disclaimers on any performance benchmark. It’s a complex topic where details matter. It’s best if you make your own picture for your own use case. We try to give a fair “arena” with our open source benchmarks, so it could be a starting point for you.
Feedback and Outlook: On-device vector search Benchmarks coming soon
We’re still working on a lot of stuff (as always 😉 and with on-device / local vector search being a completely new technology for Android, we need your feedback, creativity and support more than ever. We’ll also soon release benchmarks on the vector search. Follow us on LinkedIn, GitHub, or Twitter to keep up-to-date.