UBS unlocks advanced AI techniques with PostgreSQL on Azure

This blog was authored by Jay Yang, Executive Director, and Orhun Oezbek, GenAI Architect, UBS RiskLab 

UBS Group AG is a multinational investment bank and world-leading asset manager that manages $5.7 trillion in assets across 15 different markets. We continue to evolve our tools to suit the needs of data scientists and to integrate the use of AI. Our UBS RiskLab data science platform helps over 1,200 UBS data scientists expedite development and deployment of their analytics and AI solutions, which support functions such as risk, compliance, and finance, as well as front-office divisions such as investment banking and wealth management. RiskLab and UBS GOTO (Group Operations and Technology Office) have a long-term AI strategy to provide a scalable and easy-to-use AI platform. This strategy aims to remove friction and pain points for users, such as developers and data scientists, by introducing DevOps automation, centralized governance and AI service simplification. These efforts have significantly democratized AI development for our business users. 

This blog walks through how we created two RiskLab products using Azure services. We also explain how we’re using Azure Database for PostgreSQL to power advanced Retrieval Augmented-Generation (RAG) techniques—such as new vector search algorithms, parameter tuning, hybrid search, semantic ranking, and a graphRAG approach—to further the work of our financial generative AI use cases. 

• The RiskLab AI Common Ecosystem (AICE) provides fully governed and simplified generative AI platform services, including: 

• • Governed production data access for AI development 

• • Managed large language model (LLM) endpoints access control  

• • Tenanted RAG environments 

• • Enhanced document insight AI processing 

• • Streamlined AI agent standardization, development, registration, and deployment solutions 

• • End-to-end machine learning (ML) model continuous integration, training, deployment, and monitoring processes  

• The AICE Vector Embedding Governance Application (VEGA) is a fully governed and multi-tenant vector store built on top of Azure Database for PostgreSQL that provides self-service vector store lifecycle management and advanced indexing and retrieval techniques for financial RAG use cases.  

A focus on best practices like AIOps and MLOps 

As generative AI gained traction in 2023, we noticed the need for a platform that simplified the process for our data scientists to build, test, and deploy generative AI applications. In this age of AI, the focus should be on data science best practices—GenAIOps and MLOps. Most of our data scientists aren’t fully trained on MLOps, GenAIOps, and setting up complex pipelines, so AICE was designed to provide automated, self-serve DevOps provisioning of the Azure resources they need, as well as simplified MLOps and AIOps pipelines libraries. This removes operational complexities from their workflows.  

The second reason for AICE was to make sure our data scientists were working in fully governed environments that comply with data privacy regulations from the multiple countries in which UBS operates. To meet that need, AICE provides a set of generative AI libraries that fully manages data governance and reduces complexity. 

Overall, AICE greatly simplifies the work for our data scientists. For instance, the platform provides managed Azure LLM endpoints, MLflow for generative AI evaluation, and AI agent deployment pipelines along with their corresponding Python libraries. Without going into the nitty gritty of setting up a new Azure subscription, managing MLFlow instances, and navigating Azure Kubernetes Service (AKS) deployments, data scientists can just write three lines of code to obtain a fully governed and secure generative AI ecosystem to manage their entire application lifecycle.  

And, as a governed, secure lab environment, they can also develop and prototype ML models and generative AI applications in the production tier. We found that providing production read-only datasets to build these models significantly expedites our AI development. In fact, the process for developing an ML model, building a pipeline for model training, and putting it into production has dropped from six months to just one month.  

Azure Database for PostgreSQL and pgvector: The best of both worlds for relational and vector databases 

Once AICE adoption ramped up, our next step was to develop a comprehensive, flexible vector store that would simplify vector store resource provisioning while supporting hundreds of RAG use cases and tenants across both lab and production environments. Essentially, we needed to create RAG as a Service (RaaS) so our data scientists could build custom AI solutions in a self-service manner. 

When we started building VEGA and this vector store, we anticipated that effective RAG would require a diverse range of search capabilities covering not only vector searches but also more traditional document searches or even relational queries. Therefore, we needed a database that could pivot easily. We were looking for a really flexible relational database and decided on Azure Database for PostgreSQL. 

For a while, Azure Database for PostgreSQL has been our go-to database at RiskLab for our structured data use cases because it’s like the Swiss Army Knife of databases. It’s very compact and flexible, and we have all the tools we need in a single package. Azure Database for PostgreSQL offers excellent relational queries and JSONB document search. When used in conjunction with the pgvector extension for vector search, we created some very powerful hybrid search and hierarchical search RAG functionalities for our end users. 

The relational nature of Azure Database for PostgreSQL also allowed us to build a highly regulated authorization and authentication mechanism that makes it easy and secure for data scientists to share their embeddings. This involved meeting very stringent access control policies so that users’ access to vector stores is on a need-to-know basis. Integrations with the Azure Graph API help us manage those identities and ensure that the environment is fully secure. Using VEGA, data scientists can just click a button to add a user or group and provide access to all their embeddings/documents. It’s very easy, but it’s also governed and highly regulated.  

Speeding vector store initialization from days to seconds 

With VEGA, the time it takes to provision a vector store has dropped from days to less than 30 seconds. Instead of waiting days on a request for new instances of Azure Database for PostgreSQL, pgvector, and Azure AI Search, data scientists can now simply write five lines of code to stand up virtual, fully governed, and secure collections. And the same is true for agentic deployment frameworks. This speed is critical for lab work that involves fast iterations and experiments. And because we built on Azure Database for PostgreSQL, a single instance of VEGA can support thousands of vector stores. It’s cost-effective and seamlessly scales. 

Creating a hybrid search to analyze thousands of documents 

Since launching VEGA, one of the top hybrid search use cases has been Augmented Indexing Search (AIR Search), allowing data scientists to comb through financial documents and pinpoint the correct sections and text. This search uses LLMs as agents that first filter based on metadata stored in JSONB columns of the Azure Database for PostgreSQL, then apply vector similarity retrieval. Our thousands of well-structured financial documents are built with hierarchical headers that act as metadata, providing a filtering mechanism for agents and allowing them to retrieve sections in our documents to find precisely what they’re looking for. Because these agents are autonomous, they can decide on the best tools to use for the situation—either metadata filtering or vector similarity search. As a hybrid search, this approach also minimizes AI hallucinations because it gives the agents more context to work with. 

To enable this search, we used ChatGPT and Azure OpenAI. But because most of our financial documents are saved as PDFs, the challenge was retaining hierarchical information from headers that were lost when simply dumping in text from PDFs. We also had to determine how to make sure ChatGPT understood the meaning behind aspects like tables and figures. As a solution, we created PNG images of PDF pages and told ChatGPT to semantically chunk documents by titles and headers. And if it came across a table, we asked it to provide a YAML or JSON representation of it. We also asked ChatGPT to interpret figures to extract information, which is an important step because many of our documents contain financial graphs and charts. 

We’re now using Azure AI Document Intelligence for layout detection and section detection as the first step, which simplified our document ingestion pipelines significantly. 

Forecasting economic implications with PostgreSQL Graph Extension 

Since creating AICE and VEGA using Azure services, we’ve significantly enhanced our data science workflows. We’ve made it faster and easier to develop generative AI applications thanks to the speed and flexibility of Azure Database for PostgreSQL. Making advanced AI features accessible to our data scientists has accelerated innovation in RiskLab and ultimately allowed UBS to deliver exceptional value to our customers. 

Looking ahead, we plan to use the Apache AGE graph extension in Azure Database for PostgreSQL for macroeconomics knowledge retention capabilities. Specifically, we’re considering Azure tooling such as GraphRAG to equip UBS economist and portfolio managers with advanced RAG capabilities. This will allow them to retrieve more coherent RAG search results for use cases such as economics scenario generation and impact analysis, as well as investment forecasting and decision-making.  

For instance, a UBS business user will be able to ask an AI agent: if a country’s interest rate increases by a certain percentage, what are the implications to my client’s investment portfolio? The agent can perform a graph search to obtain all other connected economic entity nodes that might be affected by the interest rate entity node in the graph. We anticipate the AI-assisted graph knowledge will gain significant traction in the financial industry. 

Learn more 

For a deeper dive on how we created AICE and VEGA, check out this on-demand session from Ignite. We talk through our use of Azure Database for PostgreSQL and pgvector, plus we show a demo of our GraphRAG capabilities. 

About Azure Database for PostgreSQL 

Azure Database for PostgreSQL is a fully managed, scalable, and secure relational database service that supports open-source PostgreSQL. It enables organizations to build and manage mission-critical applications with high availability, built-in security, and automated maintenance.