On-device Vector Database for Dart/Flutter

On-device Vector Database for Dart/Flutter

ObjectBox 4.0 introduces the first on-device vector database for the Dart/Flutter platform, allowing Dart developers to enhance their apps with AI in ways previously not possible. A vector database facilitates advanced data processing and analysis, such as measuring semantic similarities across different document types like images, audio files, and texts. If you want to go all-in with on-device AI, combine the vector search with a large language model (LLM) and make the two interact with individual documents. You may have heard of this as “retrieval-augmented generation” (RAG). This is your chance to explore this as one of the first Dart developers.

Vector Search for Dart/Flutter

Now, let’s look into the Dart specifics! 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 highly scalable and can find relevant data within millions of entries in a matter of milliseconds.

Let’s have a deeper look into the example used in our vector search documentation. In this example, 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 (it would also need additional properties like an ID and a name, of course):

Vector objects are inserted as usual (the indexing is done automatically behind the scenes):

To perform a nearest neighbor search, use the new nearestNeighborsF32(queryVector, maxResultCount) query condition and the new “find with scores” query methods (the score is the distance to the query vector). For example, to find the 2 closest cities:

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 vectors from?

For most AI applications, vectors are created by a so-called 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. Just set up an account at your favorite AI provider and create embeddings online. Alternatively, you can also run your embedding model locally on device. This might require some research. A good starting point for that may be TensorFlow lite, which also has a Flutter package. If you want to use really good embedding models (starting at around 90 MB), you can also check these on-device embedding models. These might require a more capable inference runtime though. E.g. if you are targeting desktops, you could use ollama (e.g. using this package).

CRUD benchmarks 2024

A new release is also a good occasion to refresh our open source benchmarks. Have a look:

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.

What’s next?

We are excited to see what you will build with the new vector search. Let us know! And please give us feedback. It’s the very first release of an on-device vector database ever – and the more feedback we get on it, the better the next version will be.

The first On-Device Vector Database: ObjectBox 4.0

The first On-Device Vector Database: ObjectBox 4.0

The new on-device vector database enables advanced AI applications on small restricted devices like mobile phones, Raspberry Pis, medical equipment, IoT gadgets and all the smart things around you. It is the missing piece to a fully local AI stack and the key technology to enable AI language models to interact with user specific data like text and images without an Internet connection and cloud services.

An AI Technology Enabler

Recent AI language models (LLMs) demonstrated impressive capabilities while being small enough to run on e.g. mobile phones. Recent examples include Gemma, Phi3 and OpenELM. The next logical step from here is to use these LLMs for advanced AI applications that go beyond a mere chat. A new generation of apps is currently evolving. These apps create “flows” with user specific data and multiple queries to the LLM to perform complex tasks. This is also known as RAG (retrieval augmented generation), which, in its simplest form, allows one to chat with your documents. And now, for the very first time, this will be possible to do locally on restricted devices using a fully fledged embedded database.

What is special about ObjectBox Vector Search?

We know restricted devices. Where others see limitations, we see the potential and we have repeatedly demonstrated creating superefficient software for these. And thus maximizing speed, minimizing resource use, saving battery life and CO2. With this knowledge, we approached vector search in a unique way.

Efficient memory management is the key. The challenge with vector data is that on the one hand, it consumes a lot of memory – while on the other hand, relevant vectors must be present in memory to compute distances between vectors efficiently. For this, we introduced a special multi-layered caching that gives the best performance for the full range of devices; from memory-constrained small devices to large machines that can keep millions of vectors in memory. This worked out so well that we saw ObjectBox outperform several vector databases built for servers (open source benchmarks coming soon). This is no small feat given that ObjectBox still holds up full ACID properties, e.g. caching must be transaction-aware.

Also, keep in mind that ObjectBox is a fully capable database that allows you to store complex data objects along with vectors. From an ObjectBox data model point of view, a vector is “just” another property type. This allows you to store all your data (vectors along with objects) in a single database. This “one database” approach also includes queries. You can already combine vector search with other conditions. Note that some limitations still apply with this initial release. Full hybrid search is close to being finished and will be part of one of the next releases.

In short, the following features make ObjectBox a unique vector database:

  • Embedded Database that runs inside your application without latency
  • Vector search based is state-of-the-art HNSW algorithm that scales very well with growing data volume
  • HNSW is tightly integrated within our internal database. Vector Search doesn’t just run “on top of database persistence”.
  • With this deep integration we do not need to keep all vectors in memory.
  • Multi-layered caching: if a vector is not in-memory, ObjectBox fetches it from disk.
  • Not just a vector database: you can store any data in ObjectBox, not just vectors. You won’t need a second database.
  • Low minimum hardware requirements: e.g. an old Raspberry Pi comfortably runs ObjectBox smoothly.
  • Low memory footprint: ObjectBox itself just takes a few MB of memory. The entire binary is only about 3 MB (compressed around 1 MB).
  • Scales with hardware: efficient resource usage is also an advantage when running on more capable devices like the latest phones, desktops and servers.
  • ObjectBox additionally offers commercial editions, e.g. a Server Cluster mode, GraphQL, and of course, ObjectBox Sync, our data synchronization solution.

Why is this relevant? AI anywhere & anyplace

With history repeating itself, we think AI is in a “mainframe era” today. Just like clunky computers from decades before, AI is restricted to big and very expensive machines running far away from the user. In the future, AI will become decentralized, shifting to the user and their local devices. To support this shift, we created the ObjectBox vector database. Our vision is a future where AI can assist everyone, anytime, and anywhere, with efficiency, privacy, and sustainability at its core.

What do we launch today?

Today, we are releasing ObjectBox 4.0 with Vector Search for a variety of languages:

*) We acknowledge Python’s popularity within the AI community and thus have invested significantly in our Python binding over the last months to make it part of this initial release. Since we still want to smooth out some rough edges with Python, we decided to label Python an alpha release. Expect Python to quickly catch up and match the comfort of our more established language bindings soon (e.g. automatic ID and model handling).

Let’s get you started right away? Check our Vector Search documentation to see how to use it!

One more thing: ObjectBox Open Source Database (OSS)

We are also very happy to announce that we will fully open source the core of ObjectBox. As a company we follow the open core model. Since we still have some cleaning up to do, this will happen in one of the next releases, likely 4.1.

“Release week”

With today’s initial releases, we are far from done yet. Starting next Tuesday, you can  expect additional announcements from us. Follow us to get the news as soon as it is released.

What’s next?

This is our very first version of a “vector database”. And while we are very happy with this release, there are still so many things to do! For example, we will optimize vector search by adding vector quantization and integrate it more tightly with our data synchronization. We are also focusing on expanding our solution’s reach through strategic partnerships. If you think you are a good fit, let us know. And as always, we are very eager to get some feedback from you! Take care.

ObjectBox Database Java / Kotlin 3.0 + CRUD Benchmarks

ObjectBox Database Java / Kotlin 3.0 + CRUD Benchmarks

The Android database for superfast Java / Kotlin data persistence goes 3.0. Since our first 1.0-release in 2017 (Android-first, Java), we have released C/C++, Go, Flutter/Dart, Swift bindings, as well as Data Sync and we’re thrilled that ObjectBox has been used by over 800,000 developers.

We love our Java / Kotlin community ❤️ who have been with us since day one. So, with today’s post, we’re excited to share a feature-packed new major release for Java Database alongside CRUD performance benchmarks for MongoDB Realm, Room (SQLite) and ObjectBox.

What is ObjectBox?

ObjectBox is a high performance database and an alternative to SQLite and Room. ObjectBox empowers developers to persist objects locally on Mobile and IoT devices. It’s a NoSQL ACID-compliant object database with an out-of-the-box Data Sync providing fast and easy access to decentralized edge data (Early Access).

New Query API

A new Query API is available that works similar to our existing Dart/Flutter Query API and makes it easier to create nested conditions:

In Kotlin, the condition methods are also available as infix functions. This can help make queries easier to read:

Unique on conflict replace strategy

One unique property in an @Entity can now be configured to replace the object in case of a conflict (“onConflict”) when putting a new object.

This can be helpful when updating existing data with a unique ID different from the ObjectBox ID. E.g. assume an app that downloads a list of playlists where each has a modifiable title (e.g. “My Jam”) and a unique String ID (“playlist-1”). When downloading an updated version of the playlists, e.g. if the title of “playlist-1” has changed to “Old Jam”, it is now possible to just do a single put with the new data. The existing object for “playlist-1” is then deleted and replaced by the new version.

Built-in string array and map support

String array or string map properties are now supported as property types out-of-the-box. For string array properties it is now also possible to find objects where the array contains a specific item using the new containsElement condition.

Kotlin Flow, Android 12 and more

Kotlin extension functions were added to obtain a Flow from a BoxStore or Query:

Data Browser has added support for apps targeting Android 12.

For details on all changes, please check the ObjectBox for Java changelog.

Room (SQLite), Realm & ObjectBox CRUD performance benchmarks

We compared against the Android databases, MongoDB Realm and Room (on top of SQLite) and are happy to share that ObjectBox is still faster across all four major database operations: Create, Read, Update, Delete.

Android database comparative benchmarks for ObjectBox, Realm, and Room

We benchmarked ObjectBox along with Room 2.3.0 using SQLite 3.22.0 and MongoDB Realm 10.6.1 on an Samsung Galaxy S9+ (Exynos) mobile phone with Android 10. All benchmarks were run 10+ times and no outliers were discovered, so we used the average for the results graph above. Find our open source benchmarking code on GitHub and as always: feel free to check them out yourself. More to come soon, follow us on Twitter or sign up to our newsletter to stay tuned (no spam ever!).

Using a fast on-device database matters

A fast local database is more than just a “nice-to-have.” It saves device resources, so you have more resources (CPU, Memory, battery) left for other resource-heavy operations. Also, a faster database allows you to keep more data locally with the device and user, thus improving privacy and data ownership by design. Keeping data locally and reducing data transferal volumes also has a significant impact on sustainability.

Sustainable Data Sync

Some data, however, you might want or need to synchronize to a backend. Reducing overhead and synchronizing data selectively, differentially, and efficiently reduces bandwidth strain, resource consumption, and cloud / Mobile Network usage – lowering the CO2 emissions too. Check out ObjectBox Data Sync, if you are interested in an out-of-the-box solution.

Get Started with ObjectBox for Java / Kotlin Today

ObjectBox is free to use and you can get started right now via this getting-started article, or follow this video.

Already an ObjectBox Android database user and ready to take your application to the next level? Check out ObjectBox Data Sync, which solves data synchronization for edge devices, out-of-the-box. It’s incredibly efficient and (you guessed it) superfast 😎

We ❤️ your Feedback

We believe, ObjectBox is super easy to use. We are on a mission to make developers’ lives better, by building developer tools that are intuitive and fun to code with. Now it’s your turn: let us know what you love, what you don’t, what do you want to see next? Share your feedback with us, or check out GitHub and up-vote the features you’d like to see next in ObjectBox.

ObjectBox Dart/Flutter v0.11 Database: Performance & Relations

ObjectBox Dart/Flutter v0.11 Database: Performance & Relations

Flutter Databases are few. Therefore, we’re happy to take a big step towards 1.0 with this ObjectBox Dart v0.11 release, improving performance and bringing the much-desired relations support known from other ObjectBox DB language bindings to Dart/Flutter.

For those of you new to ObjectBox: ObjectBox is a superfast NoSQL object database for Flutter / Dart and here is how you can save data in your Dart / Flutter apps:

To learn about more ObjectBox features, like relations, queries and data sync, check our ObjectBox pub.dev page.

How fast is ObjectBox Dart? Performance Benchmarks

Speed is important for data persistence solutions. Accordingly, we wanted to test how ObjectBox compares performance-wise to other Flutter Dart database options. Therefore, we looked for libraries with comparable levels of storage abstraction and feature set –  so not just plain SQL/Key-value storage but also ORM-like features. There doesn’t seem to be that much choice… 

We looked at some two popular approaches: sqflite a SQLite wrapper for Flutter (no Dart Native support), and Hive, a key-value store with Class-adapters which seems still popular although its technology is phased out (see below). As a third alternative we pulled in Firestore, which does not really fit as it is no local database, but would be fun to compare anyway.

What we tested

To get an overview of the databases, we tested CRUD operations (create, read, update, delete). Each test was run multiple times and executed manually outside of the measured time. Data preparation and evaluation was also done outside of the measured time.

We tried to keep the test implementations as close as possible to each other while picking the approaches recommended by the docs for each database. We open sourced the test code at https://github.com/objectbox/objectbox-dart-performance if you want to have a closer look.

Performance Benchmark Results

Looking at the results, we can see ObjectBox performing significantly faster than sqflite across the board, with up to 70 times speedup in case of create & update operations. Compared to Hive, the results are a little more varied, with Hive being faster at reading objects than ObjectBox (we come to that later in our outlook), and ObjectBox being faster at creating objects, about four times faster at updates and three times for deletes. As a mostly-online database, it becomes clear that Firestore’s performance is not really comparable.

Implementation notes

ObjectBox: This release largely boosted the performance. The remaining bottlenecks are due to Dart itself and how it allows to modify byte buffers. There’s potential to double the speed if we look at other languages supported by ObjectBox. And if that’s not happening soon, we’d still have the option to do some low-level hacks…

Sqflite: a wrapper around SQLite, which is a relational database without direct support for Dart objects. Each dart object field is mapped to a column in the database, as per sqflite docs, i.e. converting between the Dart class and a Map.

Hive: We’ve tested with the latest Hive release, which is technically discontinued. The author hit two architectural roadblocks (RAM usage and queries) and is currently in the process to do a rewrite from scratch.
Update: strictly speaking it’s not straightforward to directly compare e.g. ObjectBox vs. Hive. In Hive, the high read numbers result from Dart objects already cached in memory. If the objects are fetched using the async API from disk, the numbers drop by factor 1000.

Firestore: This is totally apples and oranges, but we still decided to include Firebase/Firestore as it seems at least somewhat popular to “persist data”. It’s quite Cloud centric and thus offers limited offline features. For example, in order to use batches (“transactions”), an internet connection is required to “commit”. Also, due to its low performance, the test configuration was different: batches of 500 objects and only 10 runs.

Test setup

We ran the benchmarks as a Flutter app on a Android 10 device with a Kirin 980 CPU. The app executed all operations in batches of 10.000 objects, with each batch forming a single transaction. Each test was run 50 times, averaging the results over all the runs. This ensured the VM warmup (optimization) during the first run and garbage collections don’t affect the overall result significantly. (We care about accurate benchmarks; read more about our benchmarking best practices here.)

Outlook

With this latest release, we’re not far away from a stable API for a 1.0 release (🎉), so please share your thoughts and feedback. For the next release, we’ll add features like async operations, more relation types and some smaller improvements. We are also working on an ObjectBox variant for the Web platform that is planned close to the 1.0 release. And of course there is ObjectBox Data Sync for Flutter/Dart. If you want to be first in line to try, drop us a line, we can put you on the shortlist.

ObjectBox Swift 1.4 – In Relation to…

ObjectBox Swift 1.4 – In Relation to…

ObjectBox for Swift 1.4 makes object relations more natural and intuitive for Swift developers. For example, let’s take the teacher-student relation to Swift and how you store objects in the database. Let’s say “Teacher” is a Swift class that has a collection called “students”. Now let’s say we have a new teacher with new students and want to store them in the ObjectBox database. It’s done like this:

let yoda = Teacher(name: "Yoda")
yoda.students.append(Student(name: "Luke"))
yoda.students.append(Student(name: "Anakin"))
try box.put(yoda)

This is pretty much standard Swift. A single put command is enough to store all three new objects in the database (sorry for the “try”, Yoda, but you know, IO…). Now let’s see how this works. The students’ property in the Teacher’s class is of type ToMany<Student> and works like any Swift collection. This is because ToMany implements the protocols RandomAccessCollection and RangeReplaceableCollection. Under the hood however, it tracks all changes. Thus, when ObjectBox is instructed to put Yoda in a box, it also knows that two students were added. It also knows that our two Jedi students are new and thus puts them in database too. If you supply students that have been already persisted, it won’t put them. You can also mix new and existing objects.

Version 1.4 does not only bring TooMany (sic) improvements, but also brings a couple of new features, e.g. a bulk-get and read-only stores. You also may have heard of Sync (some kind of teleportation for objects, my young padawan). We’re still working on that, but we started to expose the Sync API with this release. It doesn’t come with any (space consuming) implementation so it’s really about getting early awareness and feedback. A full changelog is available at the docs.

So, time to start your (cocoa) pod again and let us know what you think. May the for… um, OK, that’s getting too many references for one article. One to many.