Build up-to-date generative AI applications with real-time vector embedding blueprints for Amazon MSK

Businesses today heavily rely on advanced technology to boost customer engagement and streamline operations. Generative AI, particularly through the use of large language models (LLMs), has become a focal point for creating intelligent applications that deliver personalized experiences. However, static pre-trained models often struggle to provide accurate and up-to-date responses without real-time data.

To help address this, we’re introducing a real-time vector embedding blueprint, which simplifies building real-time AI applications by automatically generating vector embeddings using Amazon Bedrock from streaming data in Amazon Managed Streaming for Apache Kafka (Amazon MSK) and indexing them in Amazon OpenSearch Service.

In this post, we discuss the importance of real-time data for generative AI applications, typical architectural patterns for building Retrieval Augmented Generation (RAG) capabilities, and how to use real-time vector embedding blueprints for Amazon MSK to simplify your RAG architecture. We cover the key components required to ingest streaming data, generate vector embeddings, and store them in a vector database. This will enable RAG capabilities for your generative AI models.

The importance of real-time data with generative AI

The potential applications of generative AI extend well beyond chatbots, encompassing various scenarios such as content generation, personalized marketing, and data analysis. For example, businesses can use generative AI for sentiment analysis of customer reviews, transforming vast amounts of feedback into actionable insights. In a world where businesses continuously generate data—from Internet of Things (IoT) devices to application logs—the ability to process this data swiftly and accurately is paramount.

Traditional large language models (LLMs) are trained on vast datasets but are often limited by their reliance on static information. As a result, they can generate outdated or irrelevant responses, leading to user frustration. This limitation highlights the importance of integrating real-time data streams into AI applications. Generative AI applications need contextually rich, up-to-date information to make sure they provide accurate, reliable, and meaningful responses to end users. Without access to the latest data, these models risk delivering suboptimal outputs that fail to meet user needs. Using real-time data streams is crucial for powering next-generation generative AI applications.

Retrieval Augmented Generation

Retrieval Augmented Generation (RAG) is the process of optimizing the output of an LLM so it references an authoritative knowledge base outside of its training data sources before generating a response. LLMs are trained on vast volumes of data and use billions of parameters to generate original output for tasks such as answering questions, translating languages, and completing sentences. RAG extends the already powerful capabilities of LLMs to specific domains or an organization’s internal knowledge base, all without the need to retrain the model. It’s a cost-effective approach to improving LLM output so it remains relevant, accurate, and useful in various contexts.

At the core of RAG is the ability to fetch the most relevant information from a continuously updated vector database. Vector embeddings are numerical representations that capture the relationships and meanings of words, sentences, and other data types. They enable more nuanced and effective semantic searches than traditional keyword-based systems. By converting data into vector embeddings, organizations can build robust retrieval mechanisms that enhance the output of LLMs.

At the time of writing, many processes for creating and managing vector embeddings occur in batch mode. This approach can lead to stale data in the vector database, diminishing the effectiveness of RAG applications and the responses that AI applications generate. A streaming engine capable of invoking embedding models and writing directly to a vector database can help maintain an up-to-date RAG vector database. This helps make sure generative AI models can fetch the more relevant information in real time, providing timely and more contextually accurate outputs.

Solution overview

To build an efficient real-time generative AI application, we can divide the flow of the application into two main parts:

Data ingestion – This involves ingesting data from streaming sources, converting it to vector embeddings, and storing them in a vector database
Insights retrieval – This involves invoking an LLM with user queries to retrieve insights, employing the RAG technique

Data ingestion

The following diagram outlines the data ingestion flow.

The workflow includes the following steps:

The application processes feeds from streaming sources such as social media platforms, Amazon Kinesis Data Streams, or Amazon MSK.
The incoming data is converted to vector embeddings in real time.
The vector embeddings are stored in a vector database for subsequent retrieval.

Data is ingested from a streaming source (for example, social media feeds) and processed using an Amazon Managed Service for Apache Flink application. Apache Flink is an open source stream processing framework that provides powerful streaming capabilities, enabling real-time processing, stateful computations, fault tolerance, high throughput, and low latency. It processes the streaming data, performs deduplication, and invokes an embedding model to create vector embeddings.

After the text data is converted into vectors, these embeddings are persisted in an OpenSearch Service domain, serving as a vector database. Unlike traditional relational databases, where data is organized in rows and columns, vector databases represent data points as vectors with a fixed number of dimensions. These vectors are clustered based on similarity, allowing for efficient retrieval.

OpenSearch Service offers scalable and efficient similarity search capabilities tailored for handling large volumes of dense vector data. With features like approximate k-Nearest Neighbor (k-NN) search algorithms, dense vector support, and robust monitoring through Amazon CloudWatch, OpenSearch Service alleviates the operational overhead of managing infrastructure. This makes it a suitable solution for applications requiring fast and accurate similarity-based retrieval tasks using vector embeddings.

Insights retrieval

The following diagram illustrates the flow from the user side, where the user submits a query through the frontend and receives a response from the LLM model using the retrieved vector database documents as context.

The workflow includes the following steps:

A user submits a text query.
The text query is converted into vector embeddings using the same model used for data ingestion.
The vector embeddings are used to perform a semantic search in the vector database, retrieving related vectors and associated text.
The retrieved information, along with any previous conversation history, and the user prompt are compiled into a single prompt for the LLM.
The LLM is invoked to generate a response based on the enriched prompt.

This process helps make sure the generative AI application can use the most up-to-date context when responding to user queries, providing relevant and timely insights.

Real-time vector embedding blueprints for generative applications

To facilitate the adoption of real-time generative AI applications, we are excited to introduce real-time vector embedding blueprints. This new blueprint includes a Managed Service for Apache Flink application that receives events from an MSK cluster, processes the events, and calls Amazon Bedrock using your embedding model of choice, while storing the vectors in an OpenSearch Service cluster. This new blueprint simplifies the data ingestion piece of the architecture with a low-code approach to integrate MSK streams with OpenSearch Service and Amazon Bedrock.

Implement the solution

To use real-time data from Amazon MSK as an input for generative AI applications, you need to set up several components:

An MSK stream to provide the real-time data source
An Amazon Bedrock vector embedding model to generate embeddings from the data
An OpenSearch Service vector data store to store the generated embeddings
An application to orchestrate the data flow between these components

The real-time vector embedding blueprint packages all these components into a preconfigured solution that’s straightforward to deploy. This blueprint will generate embeddings for your real-time data, store the embeddings in an OpenSearch Service vector index, and make the data available for your generative AI applications to query and process. You can access this blueprint using either the Managed Service for Apache Flink or Amazon MSK console. To get started with this blueprint, complete the following steps:

Use an existing MSK cluster or create a new one.
Choose your preferred Amazon Bedrock embedding model and make sure you have access to the model.
Choose an existing OpenSearch Service vector index to store all embeddings or create a new vector index.
Choose Deploy blueprint.

After the Managed Service for Apache Flink blueprint is up and running, all real-time data is automatically vectorized and available for generative AI applications to process.

For the detailed setup steps, see real-time vector embedding blueprint documentation

If you want to include additional data processing steps before the creation of vector embeddings, you can use the GitHub source code for this blueprint.

The real-time vector embedding blueprint reduces the time required and the level of expertise needed to set up this data integration, so you can focus on building and improving your generative AI application.

Conclusion

By integrating streaming data ingestion, vector embeddings, and RAG techniques, organizations can enhance the capabilities of their generative AI applications. Using Amazon MSK, Managed Service for Apache Flink, and Amazon Bedrock provides a solid foundation for building applications that deliver real-time insights. The introduction of the real-time vector embedding blueprint further simplifies the development process, allowing teams to focus on innovation rather than writing custom code for integration. With just a few clicks, you can configure the blueprint to continuously generate vector embeddings using Amazon Bedrock embedding models, then index those embeddings in OpenSearch Service for your MSK data streams. This allows you to combine the context from real-time data with the powerful LLMs on Amazon Bedrock to generate accurate, up-to-date AI responses without writing custom code. You can also improve the efficiency of data retrieval using built-in support for data chunking techniques from LangChain, an open source library, supporting high-quality inputs for model ingestion.

As businesses continue to generate vast amounts of data, the ability to process this information in real time will be a crucial differentiator in today’s competitive landscape. Embracing this technology allows organizations to stay agile, responsive, and innovative, ultimately driving better customer engagement and operational efficiency. Real-time vector embedding blueprint is generally available in the US East (N. Virginia), US East (Ohio), US West (Oregon), Europe (Paris), Europe (London), Europe (Ireland) and South America (Sao Paulo) AWS Regions. Visit the Amazon MSK documentation for the list of additional Regions, which will be supported over the next few weeks.

About the authors

Francisco Morillo is a Streaming Solutions Architect at AWS. Francisco works with AWS customers, helping them design real-time analytics architectures using AWS services, supporting Amazon Managed Streaming for Apache Kafka (Amazon MSK) and Amazon Managed Service for Apache Flink.

Anusha Dasarakothapalli is a Principal Software Engineer for Amazon Managed Streaming for Apache Kafka (Amazon MSK) at AWS. She started her software engineering career with Amazon in 2015 and worked on products such as S3-Glacier and S3 Glacier Deep Archive, before transitioning to MSK in 2022. Her primary areas of focus lie in streaming technology, distributed systems, and storage.

Shakhi Hali is a Principal Product Manager for Amazon Managed Streaming for Apache Kafka (Amazon MSK) at AWS. She is passionate about helping customers generate business value from real-time data. Before joining MSK, Shakhi was a PM with Amazon S3. In her free time, Shakhi enjoys traveling, cooking, and spending time with family.

Digish Reshamwala is a Software Development Manager for Amazon Managed Streaming for Apache Kafka (Amazon MSK) at AWS. He started his career with Amazon in 2022 and worked on product such as AWS Fargate, before transitioning to MSK in 2024. Before joining AWS, Digish worked at NortonLifelLock and Symantec in engineering roles. He holds an MS degree from University of Southern California. His primary areas of focus lie in streaming technology and distributed computing.