Retrieval Augmented Generation (RAG) with vector databases: Expanding AI Capabilities

Retrieval Augmented Generation (RAG) with vector databases: Expanding AI Capabilities

What is RAG?

Retrieval Augmented Generation (RAG) is a technique to enhance the intelligence of large language models (LLMs) with reliable facts from external sources to improve their answers. Most often, the additional knowledge comes from a vector database. For example, you can add internal data from your company, the latest news or the data from your personal devices to get responses that use your context. It can truly help you like an expert instead of giving generalized answers. This technique also reduces hallucinations. 

Why RAG?

Let’s take a look at the key benefits that RAG in general offers:

  • Customization and Adaptation: RAG helps LLMs to tailor responses to specific domains or use cases by using vector databases to store and retrieve domain-specific information. It turns general intelligence into expert intelligence.
  • Contextual Relevance: By incorporating information retrieved from a large corpus of text, RAG models can generate contextually relevant responses. It improves the quality of generated responses compared to traditional generation models.
  • Accuracy and diversity: Incorporation of external information also helps to generate more informative and accurate responses and keep LLM up-to-date. This also helps to avoid repetitive or generic responses and allows for more diverse and interesting conversations.
  • Cost-effective implementation: RAG requires less task-specific training data compared to fine-tuning the foundation models. When we compare retrieval augmented generation vs fine-tuning, RAG’s ability to use external knowledge stands out. While fine-tuning requires lots of labeled data, RAG can rely on external sources. This can be particularly beneficial in scenarios where annotated training data is limited or expensive to obtain, thus, providing a cost-effective implementation. 
  • Transparency: RAG models provide transparency in their responses by explicitly indicating the source of retrieved information. This allows users to understand how the model arrived at its response and helps enhance trust in the generated output.

Therefore, RAG is suitable for applications where access to a vast amount of specialized data is necessary. For example, a customer support bot that pulls details from FAQs and generates coherent, conversational responses. Another example is an email drafting tool that fetches information about recent meetings and generates a personalized summary.

How retrieval augmented generation works

Let’s discuss the mechanics of how RAG operates with databases, covering its main stages from dataset creation to response generation (see figure).

This image has an empty alt attribute; its file name is RAG.png
Retrieval augmented generation diagram

  • DB creation: Creation of external dataset

Before the real use, the vector database should be created. The new data, that lies outside the training dataset of LLM, should be identified and added to the dataset (e.g. up-to-date information or specific information). This dataset is then transferred into vector embeddings via an AI model (embedding language models) and is stored in the vector database. 

  • DB in use: Retrieval of relevant information
    Once a query comes in, it is also transferred into a vector / embedding. It is used then to retrieve the most relevant result from the database. To achieve this, RAG uses semantic search techniques also known as vector search to understand the user’s query and/or context, retrieving contextually relevant information from a large dataset. Vector search goes beyond keyword matching and focuses on semantic relationships, improving the quality of the retrieved information and the overall performance of the RAG system in generating contextually relevant responses. 
  • DB in use: Augmentation
    At this stage, the user’s query is augmented by adding the relevant data retrieved in the previous stage. Often, only the top responses from vector search are considered as relevant data. Many databases have additional filtering techniques in place here.
  • Generation
    The augmented query is sent to the LLM to generate an accurate answer.

The Role of Long Context Windows

The rise of the new LLMs with long (1+million tokens) context windows, like Gemini 1.5, raised the discussion on whether long context windows will replace RAG. A long context window enables users to directly incorporate huge amounts of data into a query. Thus, it increases context to the LLM to improve its efficiency. 

Long context length and RAG have pros and cons, and neither will kill the other. Rather than being mutually exclusive, large context windows and RAG can be complemented. Large context windows can enhance RAG applications by expanding the margin of precision and accommodating vast amounts of data. However, the capability of the model to take a long context does not mean that it can efficiently leverage all the information. If the relevant information is located in the middle of the context window, LLM’s ability to recall it is worse than the one located in the beginning. In order to use RAG with the long context window, the reranking (e.g. Cross-Encoder) should be used. The reranking model first calculates a matching score between a given query and vectors in the database (e.g. representing documents). And then it rearranges vector search results so that the most relevant ones are prioritized.

Future Directions

While RAG offers numerous benefits, there are still opportunities for improvement. Researchers are exploring ways to enhance RAG by combining it with other techniques. These include fine-tuning (RAFT) or the long context window (in combination with reranking). Another direction of research is expanding RAG capabilities by advancing data handling (including multimodal data), evaluation methodologies, and scalability. Finally, RAG is also affected by the new advances in optimizing LLMs to run locally on restricted devices (mobile, IoT), along with the emergence of the first on-device vector database. Now, RAG can be performed directly on your mobile device, prioritizing privacy, low latency, and offline capabilities.

Vector search: making sense of search queries

Vector search: making sense of search queries

Today, finding the most valuable information to your search is more complicated than finding a needle in a haystack. Traditional search engines match keywords and favor SEO-optimized content, but what if there was a way for search engines to truly understand the meaning behind our queries? Enter vector search – a powerful technology that is transforming how we navigate information, not just for users, but also for applications performing background searches. In this article, we will discuss what vector search is and how it works.

What is a vector search and why should you care?

Example Results with a traditional search for “Simple Fruit Cake”.

Vector search, which is also known as semantic search, is a technology that improves search accuracy by understanding the meaning (semantics) of the data and relations between its parts. Unlike traditional search, vector search efficiently handles synonyms, typos, ambiguous language, and broad or fuzzy queries. This is because it focuses on meaning, not just keywords.

Imagine that you are searching for a dessert to cook during the weekend. In a traditional search engine, the “simple fruit cake” query will reveal only websites that include these keywords. However, a vector search engine is able to provide results like “apple pie in 20 minutes” or “easy summer desserts”, which capture the essence of the query and align with your desire for a straightforward dessert option, providing more valuable results to you. 

At its core, vector search uses Large Language models (LLMs), like GPT, to transform data into mathematical vectors, also known as vector embeddings

What is a vector embedding?

2D Vector Space Representation. “Easy apple pie” is close to “simple fruit cake” as they are both simple and have fruit as an ingredient. “Easy chocolate mousse” shares simplicity but does not contain fruit. “Fancy plum cake” has fruit but is not simple to make. And “extravagant chocolate mousse” does not share either simplicity or fruit as an ingredient. Thus, it is the farthest from “simple fruit cake”.

A vector or vector embedding is a numerical representation of any kind of unstructured data (e.g. texts, images, videos, audio). It captures its meaning while being easy and efficient to compute with. Think of it like this: imagine you have a collection of cake recipes. You can convert each recipe into a vector embedding, which is like a unique numerical code that represents the recipe’s characteristics (ingredients, cooking methods, flavors, etc.).

Once all the recipes are encoded into embeddings, we can perform a similarity search. This means we can compare the vectors to see how similar the recipes are. For example, the vector for an easy apple pie recipe would be close to the vector for a simple fruit cake recipe because they share similar characteristics (e.g. simplicity, fruitiness). On the other hand, the vector for an extravagant chocolate mousse cake would be farther away because it involves different ingredients and methods.

How to compare vectors?

Vector similarity is a measure of how similar two vectors are (see ep. 4 of ObjectBox Bites). There are three ways to compare vectors: Jaccard Similarity, Cosine Similarity, and L2 Distance (also known as Euclidean distance). Jaccard Similarity calculates the ratio of elements that are common to both vectors divided by the total number of elements in both vectors. Cosine Similarity calculates the cosine of the angle between two vectors. The last method is the L2 distance. It calculates the straight-line distance between two points in space represented by the vectors. This is the most frequently used method in AI applications. It is important to note that the choice of vector comparison method does not affect the mechanics of similarity search.

What is a vector database and how is it related to vector search?

A vector database is a specialized database designed to store, manage, and search vectors efficiently. This efficiency is crucial for handling large datasets and performing fast vector similarity searches. Also, with a vector database, the knowledge of AI models can be improved, adapted, and updated. Therefore, today, most AI apps use a vector database.

Imagine having an AI that knows your habits, your preferences, your health data, maybe even what’s in your fridge, and can use this knowledge to suggest recipes that fit your lifestyle and individual preferences. A standard AI model doesn’t have that data and wouldn’t learn that way, but with a vector database it can. Now, when you search for a “fruit cake recipe”, using this data, it can suggest a “simple fruit cake” without sugar if you usually prefer quick, easy, and healthy recipes, or a “fancy plum cake” if you enjoy more challenging baking projects and don’t like apples. Or, a vegan option, if you have neither milk nor eggs left in the fridge.

This technique is called Retrieval-Augmented Generation (RAG). It enhances the capabilities of LLMs with additional data (e.g. personal data, company data, fresh data) stored in a vector database.

When you query a vector database, it uses the query’s vector representation to find the nearest neighbors in the database.

Nearest Neighbor Search

How do we find the nearest neighbor to our query vector? The most straightforward approach is a brute-force search. It calculates the distance between our query vector and all other vectors in the database, one by one. Any metrics discussed in “How to compare vectors” can be used. However, this brute-force approach has a time complexity of O(N*d), where N is the number of vectors and d is the dimensionality. This becomes computationally expensive for large datasets.

Since exact nearest neighbor search can be slow for massive datasets, we often turn to approximate nearest neighbor (ANN) algorithms. These algorithms prioritize efficiency by finding neighbors that are very close (but not necessarily the absolute closest) to the query vector, significantly reducing search time. 

Continuing with the cooking assistant app example, imagine you’re searching for a “fruit cake recipe”. Assume that in our database, the real closest recipe is “simple apple pie”. With a massive database, an exact nearest neighbor search might take a long time to find the perfect match. However, an ANN algorithm can quickly find a recipe that is very similar to what you’re looking for, such as a “simple fruit cake” or a “basic apple pie”, even if it might not be the exact closest match. This efficiency ensures you get relevant and useful recipe suggestions promptly, enhancing your overall experience without a noticeable compromise in quality.

Approximate Nearest Neighbour Search

Now, let’s delve into the world of Approximate Nearest Neighbor (ANN) algorithms. The way you search for nearest neighbors depends on how the data is stored in the vector database. One of the earliest ANN algorithms, established in 1975, is called k-d trees. These trees work by recursively splitting the data space using hyperplanes, making the search process more efficient (see ep. 5 of ObjectBox Bites). However, k-d trees, like many exact nearest neighbor algorithms, suffer from the dimensionality curse. This means that as the number of dimensions (features) in your data increases, the distance between points becomes less meaningful, making searching very slow in high-dimensional spaces like those used in vector databases. 

For instance, consider simple fruit recipes. With a few features, such as cooking time and number of ingredients, finding similar recipes would be relatively straightforward. However, if we also include many other features like sweetness level, calorie count, fruit type, all specific ingredients, preparation complexity, and user ratings, the number of dimensions increases significantly. In such high-dimensional spaces, the traditional k-d tree method becomes inefficient because the distances between points (recipes) become less distinct and meaningful.

To overcome this challenge, ANN algorithms leverage two main approaches: indexing methods and sketching methods. Indexing methods work by creating a hierarchical data structure that allows for faster exploration of the search space. Imagine a well-organized library with categorized sections instead of just randomly placed books.  Sketching methods, on the other hand,  don’t search the entire dataset directly.  Instead, they create compressed versions (sketches) of the data that are faster to compare with the query vector. This reduces the search time significantly. Often, these two approaches are combined for optimal performance.

A popular example of an ANN search implementation for high-dimensional data is the Hierarchical Navigable Small World (HNSW) algorithm (e.g. implemented in Azure AI). HNSW relies on graph-based indexing to efficiently navigate the data space and find nearest neighbors. For more details watch episodes 6, 7, and 8 of ObjectBox Bites miniseries, where we describe the fundamentals of HNSW.

Take-away notes

To sum up, vector search offers a significant leap forward in how we search for information. By understanding the meaning and relationships behind data, it delivers more relevant and accurate results, even for unstructured data and complex queries. This technology has the potential to revolutionize various fields, from enhancing search engines to empowering AI applications. As vector search continues to evolve, we can expect even more exciting possibilities for navigating the ever-growing ocean of information and unlocking its full potential. This includes operating with data directly on the devices it was created on, reducing cloud costs, eliminating the reliance on an internet connection, and opening up using your private data without it ever being shared (100% private). If you’re interested in other AI and vector database-related topics, check out the ObjectBox mini-series. Stay tuned for more articles in the future.

Python on-device Vector and Object Database for Local AI

Python on-device Vector and Object Database for Local AI

Python developers can now use the very first on-device object/vector database for AI applications that run everywhere, locally. With its latest release, the battle-tested ObjectBox database has extended its Python support. This embedded database conveniently stores and manages Python objects and vectors, offering highly performant vector search alongside CRUD operations for objects.

What is ObjectBox?

ObjectBox is a lightweight embedded database for objects and vectors. Note that “objects” here refers to programming language objects, e.g. instances of a Python class. And because it was built for this purpose, ObjectBox is typically the fastest database option in this category. In terms of performance, it easily beats wrappers and ORMs running on top of SQL databases. This is because middle layers like SQL and row/column mapping simply do not exist.

ObjectBox is also a vector database storing high dimensional vector data and offering a highly scalable vector search algorithm (HNSW). Even with millions of documents, ObjectBox is capable of finding nearest neighbors within milliseconds on commodity hardware. And for ObjectBox, a vector is “just another” property type and thus, you can combine vector data with regular data using your own data model.

The ObjectBox API

Note: for an interactive version of the example, check our vector search Jupyter notebook on Google Colab, or one of the two vector-search-city examples in our repository.

Having an easy-to-use API is a top priority for ObjectBox. The following example uses a City entity, which has a name and a location. The latter is a two dimensional vector of latitude and longitude. We create a Store (aka the database) with default options, and use a Box to insert a list of Cities:

With cities stored in the database, let’s do a simple search for cities starting with “Be”:

Vector search follows the same pattern. This query locates the nearest neighbors to a given location:

LangChain Integration

ObjectBox is integrated as a Vector Database in LangChain via the langchain-objectbox package:

pip install langchain-objectbox --upgrade

Then, create an ObjectBox VectorStore using e.g. one of the from_ class methods e.g. from_texts class method:

from langchain_objectbox.vectorstores import ObjectBox
obx_vectorstore = ObjectBox.from_texts(texts, embeddings, ...)

We will look into details in one of our next blog posts.

Vector Search Performance

While ObjectBox is a small database, you can expect great performance. We ran a quick benchmark on using the popular and independent ANN benchmark open source suite. First results indicate that ObjectBox’ vector search is quite fast and that it can even compete with vector databases built for servers and the cloud. For more details, we will have a special ANN benchmark post that goes in more detail soon (follow us to stay up-to-date: LinkedIn, Twitter).

From Zero to 4: our first stable Python Release

We jumped directly to version 4.0 to align with our “core” version. The core of ObjectBox is written in high-performance C++ and with the release of vector search, we updated its version to 4.0. Thus you already get all the robustness you would expect from a 4.0 version of a product that has been battle tested for years. By aligning the major version, it’s also easy to tell that all ObjectBox bindings with version 4 include vector search.

What’s next?

There are a lot of features still in the queue. For example our Python binding does not support relations yet. Also we would like to do further benchmarks and performance work specific to Python. We are also open for contributions, check our GitHub repository

Evolution of search: traditional vs vector search

Evolution of search: traditional vs vector search


In today’s digital landscape, searching for information is integral to our daily lives, whether for education, research, work, or shopping. However, as the volume and complexity of data kept growing, traditional search methods faced more and more challenges in providing accurate and relevant results. That’s where vector search comes in. We’re already seeing Google changing its search engine to empower the vector search (RankBrain, BERT, Neural matching), and expecting even greater incorporation of AI tools to improve search experience. Let’s explore the differences between traditional (keyword) search and vector search to understand how these technologies are shaping our search experiences, and how this impacts the discoverability of any content you might produce.

Traditional (keyword) search

Traditional search performs exact keyword matching from user queries to the data to retrieve relevant results. For example, searching for “programming languages” with traditional search will list every source containing those words. A more advanced version can also incorporate additional rules to enhance search results, such as:

  • keyword frequency (how often the term “programming languages” is used within the result text),
  • the presence of related terms (e.g. “Java”, “Python”, “C++” versus “cooking”, “gardening”),
  • or location (results closer to your location are favored).

While this approach has served us well, it struggles with ambiguous language, synonyms as well as the impact of SEO strategies, often resulting in less accurate or less valuable search results. This can be especially frustrating for businesses who are trying to get their content seen by the right people. For example, a business that publishes a blog post about “sustainable fashion tips” might miss out on potential customers who are searching for “eco-friendly clothing recommendations” or “green clothing ideas,” simply because their keywords don’t exactly match.

Vector (semantic) search

2D Vector Space Representation. In this space, “Python” and “Java” are close to each other as well as to the “Programming language” query we are searching for because they are similar (they share high values in their features).

On the other hand, vector search takes a different approach by seeking out related objects that share similar characteristics or semantics. You can think of it as finding results based on meaning or understanding rather than just exact wording. For example, searching for “programming languages” with vector search will not only find sources mentioning those exact words but also identify specific languages like “Python” or “Java” as well as related concepts such as “coding tutorials” or“development frameworks”.

To do a vector search, first of all, the content, such as texts, images, audio files, or videos, needs to be represented as vectors/embeddings (also often called vector embeddings) by AI models. These embeddings represent data in a multidimensional vector space. Vectors capture the essence or semantics of the data they represent while remaining computationally efficient.

Once these vector representations are generated, they are basically sets of numbers, and therefore easy to compute with. For instance, instead of searching for a specific word in text, we aim to find the closest vector (from the text embeddings) to the query vector (representing the word we’re searching for). This process relies on well-established vector computing methods, such as calculating the distance between vectors or minimizing the angle between them (Cosine similarity).


Let’s now compare different aspects of searching to understand the main differences between traditional search and vector search (also summarized in Table). 

  • Search Approach
    In traditional search, the approach relies on matching keywords directly from the user query to the content. Vector search uses vector embeddings to catch the semantics of the data to perform a meaning-based approach.
  • Ambiguity handling
    Therefore, vector search shines when it comes to handling ambiguity. It is superior for handling synonyms, ambiguous language, and broad or fuzzy queries compared to traditional search. This also automatically influences the relevance of the search results.
  • Search relevance calculation
    The metrics used for search relevance calculations are different. The traditional search uses term frequency-inverse document frequency (TF-IDF) and BM25, while vector search uses Jaccard Similarity, Cosine Similarity, and L2 Distance (or Euclidean).
  • Speed and Implementation
    Traditional search is easy to implement, straightforward in usage, and fast for simple queries. Vector search may be slower for simple queries and more complicated to implement, once it comes to huge datasets. However, the implementation of approximate nearest neighbor techniques (ANN) allows to significantly speed up the process.
  • Scalability
    Continuous expansion of content, challenges the scalability of traditional search, while for vector search scalability is one of the advantages. 
  • Cost
    While traditional search may have lower computational requirements, the superior performance and accuracy of vector search often justify the investment in additional computing power. Furthermore, the computational costs for vector search can be significantly reduced with the use of ANN.
Comparison table between keyword search and vector search


In summary, both traditional search and vector search offer distinct advantages and drawbacks. Vector search excels in handling ambiguity, correcting typos, enhancing relevance, and managing extensive datasets. Traditional search remains advantageous for straightforward queries, exact matches, or smaller datasets. Historically, limited computational resources, particularly for on-device computation (i.e. Edge Computing), favored traditional search. However, the landscape is evolving rapidly with the introduction of the first edge vector database solution by ObjectBox. This innovation promises to revolutionize the scenario by optimizing vector search for devices with constrained resources, extending the benefits of semantic search to the Edge.

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.

Edge AI: The era of on-device AI

Edge AI: The era of on-device AI

AI anywhere and anytime - free from Internet dependencies & 100% private

Edge AI is an often overlooked aspect of AI’s natural evolution. It is basically the move of AI functionalities away from the cloud (or powerful server infrastructure) towards decentralized (typically less powerful) devices at the network’s edges, including on mobile phones, smartwatches, IoT devices, microcontrollers, ECUs, or simply your local computer. Or in more broadly speaking: “Edge AI” means AI that works directly on-device, “local AI”.

Therefore, Edge AI apps work independently from an internet connection, offline as well as online. So, they are ideal for low, intermittent, or no connectivity scenarios. They are reliably available, more sustainable, and – of course – way faster on-device than anything hosted in the cloud. On-device AI apps can empower realtime AI anytime and anyplace.

Edge AI is where Edge Computing meets AI

The importance of vector databases for AI applications

To enable powerful on-device AI applications, the on-device (edge) technology stack needs local vector databases. So, before diving deeper into Edge AI, we’ll dive into vector databases first. Jump this section, if you are already familiar with them.

What is a vector database?

Just as SQL databases handle data in rows and columns, graph databases manage graphs, object databases store objects, vector databases store and manage large data sets of vectors, or more precisely, vector embeddings. Because AI models work with vector embeddings, vector databases are basically the databases for AI applications. Vector databases offer a feature set of vector operations, most notably vector similarity search, that makes it easy and fast to work with vector embeddings and in conjunction with AI models.

When and why do you need a vector database? 

Given the significance of vector embeddings (vectors) for AI models, particularly Large Language Models (LLMs) and AI applications, vector databases are now integral to the AI technology stack. They can be used to:

Train AI models (e.g. ML model training, LLM training)
Vector databases manage the datasets large models are trained on. Training AI models typically entails finding patterns in large data sets. Training ML models often involves finding patterns in large datasets. Vector databases significantly speed up identifying patterns and finding relationships by enabling efficient retrieval of similar data points.

Speed up AI model / LLM responses
Vector databases use various techniques to speed up vector retrieval and similarity search, e.g. compression and filtering. They accelerate both model training and inference, thus, enhancing the performance of generative AI applications. By optimizing vector retrieval and similarity search, vector dbs can enhance the efficiency and scalability of AI applications that rely on high-dimensional data representations

Add long-term memory to AI models and LLMs
Vector databases add long term memory to AI applications in two ways: They persist the history to 1. continue on the tasks or conversation later as needed and 2. to personalize and enhance the model for better-fitting results.

Enable Multimodel Search
Vector databases serve as the backbone to jointly analyze vectors from multimodal data (text, image, audio, and video) for unified multimodal search and analytics. The use of a combination of vectors from different modalities enables a deeper understanding of the information, leading to more accurate and relevant search results.

Enhancing LLMs responses, primarily “RAG
With a vector database, you have additional knowledge to enhance the quality of a model’s responses and to decrease hallucinations; real-time updates, as well as personalized responses, become possible.

Perform Similarity Search / Semantic Retrieval
Vector databases are the heart and soul of semantic retrieval and similarity search. Vector search often works better than „full-text search“ (FTS) as it finds related objects that share the same semantics/meaning instead of matching the exact keyword. Thus, it is possible to handle synonyms, ambiguous language, as well as broad and fuzzy queries.

Cache: Reduce LLM calls
Vector databases are used to cache similar queries and responses can be used as a lookup prior to calling the LLM. This saves resources, time, and costs.

The shift to on-device computation (aka Edge Computing)

Edge Computing is in its essence a decentralized computing paradigm and based on Edge Computing, AI on decentralized devices (aka Edge AI) becomes possible. Note: In computing, we have regularly seen shifts from centralized to decentralized computing and back again.

What is Edge Computing?

Our world is decentralized. Data is produced and needed everywhere, on a myriad of distributed devices like smartphones, TVs, robots, machines, and cars – on the so-called “edge” of the network. It would not only be unsustainable, expensive, and super slow to send all this data to the cloud, but it is also literally unfeasible. So, much of this data simply stays on the device it was created on. To harness the value of this data, the distributed “Edge Computing” paradigm is employed.

When and why do you need Edge Computing? 

Edge Computing stores and processes data locally on the device it was created on, e.g. on IoT, Mobile, and other edge devices. In practice, Edge Computing often complements a cloud setup. The benefits of extending the cloud with on-device computing are:

    • Offline-capability
      Storing and computing data directly on-device allows devices to operate independently from an Internet connection, which is crucial for remote locations (e.g. oil rigs in the ocean) or applications that need to always work (e.g., while the car is in underground garages, or in remote areas).
    • Data ownership/privacy
      Cloud apps are fundamentally non-private and limit the user’s control over their own data. Edge Computing allows data to stay where it is produced, used, and where it belongs (with the user/on the edge devices). It therefore reduces data security risks, and data privacy and ownership concerns.
    • Bandwidth constraints and the cost of data transmission
      Ever growing data volumes strain bandwidth and associated network/cloud costs, even with advanced technologies like 5G/6G networks. Storing data locally in a structured way at the edge, such as in an on-device database, is necessary to unlock the power of this data. At the same time, some of this data can still be made available centrally (in the cloud or on an on-premise server), combining the best of both worlds.
    • Fast response rates and real-time data processing
      Doing the processing directly on the device is much faster than sending data to the cloud and waiting for a response (latency). With on-device data storage and processing, real-time decision making is possible.
    • Sustainability
      By reducing data overhead and unnecessary data transfers, you can cut down 60-90% of data traffic, thereby significantly reducing the CO2 footprint of an application. A welcome side effect is that this also lowers costs tremendously.

Edge AI needs on-device vector databases

Every megashift in computing is empowered by specific infrastructure software, like e.g. databases. Shifting from AI to Edge AI, we still see a notable gap: On-device support for vector data management (the typical AI data) and data synchronization capabilities (to update AI models across devices). To efficiently support Edge AI, vector databases that run locally, on edge devices, are as crucial as they are on servers today. So far, all vector databases are cloud / server databases and cannot run on restricted devices like mobile phones and microcontrollers. But moreover, they often don’t run on more capable devices like standard PCs either, or only with really bad performance. To empower everyday life AI that works anytime all around us, we therefore need a database that can run performantly on a wide variety of devices on the edge of the network.

In fact, vector databases may be even more important on the edge than they are in cloud / server environments. On the edge, the tradeoff between accuracy and performance is a much more delicate line to walk, and vector databases are a way to balance the scales.

Edge AI Vector Databases for on-device use

On-device AI: Use Cases and why they need an Edge Vector Database

Seamless AI support where it is needed most, on everyday devices and all the things around us needs an optimized local AI tech stack that runs efficiently on the devices. From private home appliences to on-premise devices in business settings, medical equipment in healthcare, digital infrastructure in urban environments, or just mobile phones, you name it: To empower these devices with advanced AI applications, you need local vector databases. From the broad scope of AI’s impact in various fields, let’s focus on some specific examples to make it more tangible: the integration of AI within vehicle onboard systems and the use of Edge AI in healthcare.   

Vehicle onboard AI and edge vector databases – examples

Imagine a car crashing because the car software was waiting on the cloud to respond – unthinkable. The car is therefore one of the most obvious use cases for on-device AI.

Any AI application is only as good as its data. A car today is a complex distributed system on wheels, traversing a complex decentralized world. Its complexity is permanently growing due to increased data (7x more data per car generation), devices, and the number of functions. Making use of the available data inside the car and managing the distributed data flows is therefore a challenge in itself. Useful onboard AI applications depend on an on-device vector database (Edge AI). Some in-car AI application examples:

  • Advanced driver assistance systems (ADAS)
    ADAS benefit in a lot of areas from in-vehicle AI. Let’s look, for example, at driver behaviour: By monitoring the eye movements and head, ADAS can determine when the driver shows any signs of unconcentrated driving, e.g., drowsiness. Using an on-device database, the ADAS can use the historic data, the realtime data, and other car data, like, e.g., the driving situation, to deduce its action and  issue alerts, avoid collisions, or suggest other corrective measures. 
  • Personalized, next-gen driver experience
    With an on-device database and Edge AI, an onboard AI can analyze driver behavior and preferences over a longer period of time and combine it with other available data to optimize comfort and convenience for a personalised driving experience that goes way beyond a saved profile. For example, an onboard AI can adjust the onboard entertainment system continually to the driver’s detected state, the driving environment, and the personal preferences. 

Applications of Edge AI in Healthcare – examples

Edge Computing has seen massive growth in healthcare applications in the last years as it helps to maintain the privacy of patients and provides the reliability and speed needed. Artificial intelligence is also already in wide use making healthcare smarter and more accurate than ever before. With the means for Edge AI at hand, this transformation of the healthcare industry will become even more radical. With Edge AI and on-device vector databases, healthcare can rely on smart devices to react in realtime to users’ health metrics, provide personalized health recommendations, and offer assistance during emergencies – anytime and anyplace, with or without an Internet connection. And while ensuring data security, privacy, and ownership. Some examples:

  • Personalized health recommendations
    By monitoring the user’s health data and lifestyle factors (e.g. sleep hours, daily sports activity) combined with their historic medical data, if available, AI apps can help detect early signs of health issues or potential health risks for early diagnosis and intervention. The Ai app can provide personalized recommendations for exercise, diet, or medication adherence. While this case does not rely on real-time analysis and fast feedback as much as the previous example, it benefits from an edge vector database in regards to data privacy and security.
  • Point of care realtime decision support
    By deploying AI algorithms on medical devices, healthcare providers can receive immediate recommendations, treatment guidelines, and alerts based on patient-specific data. One example of where this is used with great success, is in surgeries. An operating room, today, is a complex environment with many decentralized medical devices that requires teams to process, coordinate, and act upon several information sources at one time. Ultra-low latency streaming of surgical video into AI-powered data processing workflows on-site, enables surgeons to make better informed decisions, helps them detect abnormalities earlier, and focus on the core of their task.

Edge AI: Clearing the Path for AI anywhere, anytime

For an AI-empowered world when and where needed, we still have to overcome some technical challenges. With AI moving so fast, this seems however quite close. The move into this new era of ubiqutuous AI needs Edge AI infrastructure. Only when Edge AI is so easy to implement and deploy as cloud AI, will we see the ecosystem thriving and bringing AI functionalities that work anytime and anyplace to everyone. An important corner stone will be on-device vector databases as well as new AI frameworks and models, which are specifically designed to address Edge Computing constraints. Some of the corresponding recent advances in the AI area include “LLM in a Flash” (a novel technique from Apple for effective inference of LLMs at the edge) and Liquid Neural Networks  (designed for continuous learning and adaptation on edge devices). There’s more to come, follow us to keep your edge on Edge AI News.