On EDB Postgres® AI. Q&A with Torsten Steinbach

Q1. What are the main challenges in integrating and managing AI data for critical enterprise workloads?

Modern AI application patterns are heavily leaning into real-time consumption of domain specific AI knowledge. This means that the customer’s domain specific and private data needs to be available for highly efficient retrieval any time the solution applies AI, i.e. anytime it runs an inference with an AI model. The RAG application pattern is the most prominent and popular example for this principle.

Building such a retrieval system is a genuine database problem. Consequently, you need a vector data store or AI data platform that is able to serve the retrieval queries at scale, with high reliability and response time SLAs. These AI databases also must be reliably up to date with the latest domain or private AI data. Preparing retrieval data and serving retrieval requests are effectively complex vector data pipelines that need to be automated and delivered with the highest operational qualities to support critical enterprise workloads.

Q2. What is EDB Postgres® AI? Is it built on Postgres?  

EDB Postgres AI is a platform that is built on Postgres in order to serve modern workloads, including AI solutions. It consists of a series of additional capabilities that EDB is building on and with Postgres. On one hand it addresses AI data integration, i.e. the ability to discover, connect to and consume AI data, no matter where it resides. This does of course include AI data that is already in Postgres. But it also includes support for external AI data (e.g. on object storage) as a first class citizen in Postgres. Another capability that EDB is adding to Postgres is orchestration of vector pipelines. This addresses the previously mentioned AI data preparation for retrieval and typical AI data retrieval patterns such as RAG. 

Yet another area of investment by EDB is the capability to conduct highly efficient AI search. This uses pgvector as a basis but adds various optimizations and accelerations to vector search, for instance the ability to use combined indexes for hybrid AI search across vector and relational data. Finally, EDB is also working on abilities to serve AI models right in the Postgres platform with enterprise level scalability to allow for sovereign AI solution stacks. In summary: EDB Postgres AI provides a comprehensive suite of artificial intelligence capabilities, including data integration, orchestration, search functionality, and sovereign serving, all delivered with enterprise-level quality of service.

Q3. Is your “AI database” capability a deployed product yet?

We currently have a tech preview for the AI database extension that handles orchestration of vector pipelines in Postgres. It can be downloaded and deployed as a docker image by anyone, and used in non-production settings. We will be making an announcement soon about the general availability of our AI database offering.

Q4. What does this tech preview include?

It supports AI data that is stored in Postgres already, as well as AI data that is stored on external S3-compatible object storage. It entirely automates the generation of vector embeddings for this AI data. It introduces the “retriever” as a new database asset that abstracts all this processing and serves as a mechanism to conduct the AI data search and retrieval in AI applications. The application can directly send the retrieval query to Postgres and receives the most relevant AI data in return. It doesn’t have to understand and interact with vector embeddings at all. The tech preview supports text data but also image data, including support for cross-modality search from text to image. Already, existing data in Postgres or object storage can be prepared in a bulk fashion. But for data in Postgres specifically, the tech preview also supports a real-time preparation for retrieval, which means that anytime new data is inserted to Postgres, or updated, the retriever also gets updated with this new data in real-time and there is zero delay in this data becoming available for retrieval in AI applications.

Q5.Using Postgres for AI requires processing new types of data like open table formats and binary and unstructured data like text, images, voice recordings and video data. But Postgres can’t traditionally process these types of data out of the box. How does EDB Postgres AI help manage such challenges?

We are enhancing Postgres in two significant ways. First, we’re introducing explicit functions that treat external object storage locations as first-class data citizens. Second, we’re integrating the ability to run LLMs directly within the Postgres platform. This effectively adds a new AI processing engine alongside SQL processing, enabling Postgres to natively understand multi-modal AI data.

We’re not just introducing isolated artificial intelligence capabilities; we’re offering a comprehensive, end-to-end integration platform. This unified solution encompasses data integration, orchestration, search functionality, and serving – all tailored for AI workflows. Our customers benefit from a single offering that requires only one deployment effort, providing access to these preconfigured and fully integrated capabilities. This versatile platform is available in the form factor of their choice, whether in the cloud or on-premise in a Kubernetes environment.

Q6. What do you see as the future of vectors? Will pgvector features eventually be part of Postgres, or will Postgres have its own way of doing similarity search?

As vector support gains prominence, it’s conceivable that it may eventually be integrated into Postgres core, especially given the potential for optimization and integration with Postgres’ key strengths. While EDB fully supports and will continue to embrace pgvector, we’re also developing additional differentiating features, such as hybrid search acceleration. Simultaneously, we’re actively engaging with other communities developing innovative vector extensions, like pgvecto.rs. This multi-faceted approach allows us to leverage the best of pgvector while also exploring and contributing to emerging technologies in the vector database space, ensuring we stay at the forefront of database evolution.

Q7. Anything else you wish to add?

The database landscape is evolving rapidly, with new players offering vector search-first solutions while established technologies integrate vector capabilities into their existing systems. We align with the latter approach, believing that vector search will ultimately consolidate around comprehensive database solutions. As the generative AI market matures, the focus is shifting towards sustainable, operational use and the ability to scale diverse AI solutions with reliable data management. Postgres is exceptionally well-positioned in this landscape, boasting both the most pervasive database developer community and high-end enterprise-grade operational features. This unique combination of community support and proven capabilities makes Postgres an ideal choice for organizations implementing and scaling AI solutions in production environments, particularly as the industry moves beyond the initial hype cycle towards practical, long-term AI implementations.

…………………………………………..

Torsten Steinbach, Chief Architect for Analytics & AI, EDB.

Torsten Steinbach is the Chief Architect for Analytics and AI at EDB. He is uplifting Postgres towards a platform for analytics and AI workloads, augmenting and extending it with enterprise-level Data Lakehouse and AI functionality. Torsten previously held senior positions in engineering of cloud, database and analytics technology. He has been a Distinguished Engineer at IBM and CTO for Big Data in Cloud. He has also been the lead architect for IBM’s enterprise database performance tooling as well as in-database Data Science & Machine Learning, for instance via deep embedding of Apache Spark. Torsten lives and breathes data and he is an experienced speaker in setting bold visions for database technology, data analytics, data science and AI and more in order to excite teams to pursue and deliver in complex endeavours.