Azure Database for PostgreSQL Blog

12 MIN READ

Azure PostgreSQL with Azure Open AI to innovate Banking Apps: Unlocking the Power of AI Extension

Microsoft

Oct 14, 2024

The financial sector is evolving rapidly, with artificial intelligence (AI) playing a pivotal role in transforming how banks handle critical functions. From automating customer interactions to identifying patterns in vast datasets, AI is becoming a key enabler for operational efficiency and customer satisfaction. This article delves into how financial institutions can integrate AI capabilities into their banking applications using Azure PostgreSQL Flex and Azure OpenAI, focusing on practical solutions like automating customer complaint management and enhancing data-driven decision-making.

This blog will guide you through how to use Azure PostgreSQL Flex integrated with Azure AI services, such as Azure OpenAI, to make existing applications intelligent and build a solution that automatically processes and retrieves customer complaints, provides semantic search capabilities, and delivers faster resolutions.

Why AI for Banking Applications?

Banks are uniquely positioned to benefit from AI due to the massive volumes of structured and unstructured data they generate daily. Whether it's processing customer complaints, identifying fraud, or personalizing services, AI can streamline and enhance these processes significantly. This blog outlines how Azure PostgreSQL Flex, a scalable cloud database, combined with Azure OpenAI services, can modernize existing banking applications by adding intelligent, AI-powered features.

AI Use Cases for Banking

Automated Customer Complaint Management: AI can revolutionize how banks handle customer complaints by automating categorization, generating responses for common issues, and escalating more complex cases. This not only speeds up the resolution process but also improves overall customer satisfaction.
Fraud Detection and Prevention: By leveraging machine learning models, banks can detect fraudulent transactions in real-time. These AI models can analyze patterns and anomalies in transaction data, helping prevent fraud before it impacts the customer.
Predictive Analytics for Credit Risk: AI models integrated with Azure PostgreSQL Flex can assess a customer’s creditworthiness based on historical data, enabling banks to offer personalized loans and credit terms while minimizing risk.

Key Azure AI Capabilities with Azure PostgreSQL Flex

To enable these use cases, Azure PostgreSQL Flex integrates seamlessly with several AI technologies, providing a flexible and secure backend for your banking applications:

PgVector Extension: Enables similarity search with AI-generated vector embeddings. This is particularly useful for tasks like semantic search in complaint management systems.
Generative AI & Retrieval-Augmented Generation (RAG): Integrates private data into AI responses, allowing banks to leverage domain-specific knowledge.
In-Database Embedding: Supports real-time embedding generation for low-latency applications, making it ideal for handling dynamic workloads in banking.

When combined with Azure OpenAI, it becomes possible to create intelligent systems capable of automating customer interactions, such as classifying and responding to complaints. By leveraging AI-powered services, banks can automatically categorize complaints, generate automated responses for simpler issues, and escalate more complex cases to the appropriate departments for further review and response.

This integration delivers significant benefits, including improved efficiency by reducing manual intervention, faster complaint resolution, and enhanced customer satisfaction. Additionally, the scalability of Azure PostgreSQL Flex ensures that banks can handle varying complaint volumes without performance issues, while maintaining regulatory compliance. Together, these technologies offer a powerful solution for modernizing banking applications, helping financial institutions optimize operations and better serve their customers.

But first, let's look at some of the Azure AI capabilities and how they fit seamlessly into PostgreSQL.

Azure AI Capabilities with Azure PostgreSQL Flex

Summary:

Capability	Description	Key Benefits
Pgvector Extension	Stores, indexes, and queries vectors for AI-driven similarity searches.	Multiple distance functions, seamless OLTP integration, hybrid search, secure
Generative AI & RAG	Build dynamic AI responses and integrates private data into LLM outputs.	Enriches AI with domain-specific knowledge
Azure AI Integration (azure_ai)	SQL-based interface for integrating AI services like OpenAI, NLP, Translator, and Machine Learning.	No complex re-architecture, supports advanced AI capabilities
In-Database Embedding (azure_local_ai)	Generates embeddings within PostgreSQL for low-latency and real-time workloads.	Faster creation, no external setup, data stays local
Vector Generation	Supports both remote and in-database embedding generation.	Flexible embedding models

Pgvector Extension for Seamless Integration with AI Solutions

The Pgvector extension in PostgreSQL allows the storage, indexing, and querying of vectors, which are fundamental for AI-driven similarity search scenarios. Currently, pgvector version 0.7 is supported which provides additional measurement options and better performance. This extension supports various vector distance functions, making it ideal for applications like recommendation systems and semantic search.

Key Benefits:

Multiple Vector Distance Functions: Supports different similarity measures for comparing vectors.
Seamless Integration: Allows AI-powered solutions to integrate directly into existing OLTP (Online Transaction Processing) PostgreSQL applications without needing to export data to specialized systems.
Enterprise-Ready Features: Includes access control, encryption, high availability, and disaster recovery, ensuring a secure and scalable infrastructure.
Hybrid Search: Combines vector search with row filtering and full-text search, ideal for complex query scenarios.
DiskANN Vector Index: It enables accurate and fast query results using WHERE clauses and reduces the memory footprint and cost of vector workloads by leveraging SSD storage with DiskANN. It applies compression and quantization techniques to enhance speed, accuracy, and latency for vector searches without consuming excessive resources

Integration with Generative AI development Frameworks

Generative AI applications can be built to respond to queries with dynamic responses, based on vector search results. Integrating generative AI development frameworks like LangChain with Azure PostgreSQL Flex allows for seamless query generation and advanced data retrieval by leveraging AI models for natural language processing. This integration enables the use of AI-driven tools to generate intelligent insights from complex datasets in PostgreSQL, automating tasks like query optimization, data summarization, and enhancing user interaction through conversational interfaces. RAG apps can retrieve private data and integrate it into LLM (Large Language Model) responses, enriching the AI's outputs with domain-specific knowledge stored in the database.

Key Benefits:

Enhanced Data Retrieval: AI-powered, natural language-driven queries make it easier to retrieve relevant information from large datasets.
Improved User Experience: Conversational AI interfaces allow for more intuitive interaction with data, providing users with actionable insights.
Another key benefit of RAG (Retrieval-Augmented Generation) is the improved relevance of LLM responses due to grounding in factual data. By retrieving and incorporating real-time, relevant information from external sources, RAG ensures that the generated responses are not only more contextually accurate but also based on up-to-date, factual data, enhancing the quality and reliability of the response.
Advance form of RAG aka GraphRAG can also be leveraged which utilizes LLM-generated knowledge graphs to significantly enhance question-and-answer performance, particularly when analyzing complex documents. By structuring information into a graph format, GraphRAG allows the system to identify and understand intricate relationships between entities, concepts, and data points, leading to more accurate and contextually relevant answers. This approach is especially beneficial in handling complex information, as it enables deeper insights and more effective retrieval of relevant knowledge.
Automated Insights Generation: Streamlines query optimization, data summarization, and other complex tasks, improving operational efficiency.
Domain-Specific Knowledge Integration: Enriches LLM responses with specialized data from private databases, enabling more accurate and relevant outputs for specific business needs.

Azure AI Integration via azure_ai Extension

The azure_ai extension offers a SQL-based interface to interact with AI services directly from PostgreSQL, making it easy for developers to integrate AI capabilities without the need for complex re-architecture. Azure AI extension gives the ability to invoke the Azure AI Language Services such as sentiment analysis right from within the database.

Key Benefits:

Azure OpenAI for embedding models and semantic searches.
Azure AI Language Services provide natural language processing (NLP) feature with real-time translation with support for 100+ languages, sentiment analysis, abstractive and extractive summarization.
Azure AI Translator for multilingual support.
Azure Machine Learning for more advanced predictive models.

In-Database Embedding Models

The azure_local_ai extension enables in-database embedding generation, powered by Microsoft's open-source E5 embedding model. This extension allows for low-latency embedding creation directly within PostgreSQL, ideal for real-time OLTP workloads.

Key Benefits:

Faster Creation: Embedding generation is significantly faster with local processing, making it ideal for frequently changing data.
No External Setup: There's no need for external services, reducing costs and maintenance.
Data Stays Local: All data remains within PostgreSQL, ensuring compliance and security.

Azure PostgreSQL offers the flexibility to generate embeddings either remotely using Azure OpenAI or locally within the database using azure_local_ai.

Remote:

SELECT * FROM <table>
ORDER BY
database_description <->
azure_openai.create_embeddings('text-embedding-ada-002', 'Databases with vector support');

In-Database:

SELECT * FROM <table>
ORDER BY
recipe_embedding <#>
azure_local_ai.create_embeddings('multilingual-e5-small:v1', 'Databases with vector support');

Why Use Azure PostgreSQL and AI for Banking Intelligent Applications?

Azure PostgreSQL Flex provides scalability and high availability, making it ideal for large-scale banking applications. By combining it with Azure AI, banks can leverage natural language processing (NLP) to automatically analyze, categorize, and resolve customer complaints. This not only improves complaint resolution times but also enhances customer experience by using AI to identify patterns and offer personalized solutions within their existing applications or new systems.

How to Integrate AI into Banking Applications Using Azure PostgreSQL Flex

A smart complaint management system for banking, powered by Azure OpenAI integrated with Azure PostgreSQL, revolutionizes how financial institutions handle customer complaints. This system can automate the entire process from complaint intake to resolution, providing faster, more accurate responses, enhancing customer satisfaction, and reducing operational bottlenecks. By leveraging the power of Azure OpenAI's natural language processing (NLP) models, such a system can understand and categorize customer complaints in real-time, while Azure PostgreSQL serves as the scalable, secure backend for storing and retrieving customer data, complaint history, and resolution actions.

Let's walk through a step-by-step process, including sample SQL queries and their results, to implement a smart complaint management system.

Creating a Customer Complaints Table in Azure PostgreSQL

First, we create a table to store customer complaints, capturing key information such as complaint ID, title, description, and status.

SQL Query:

-- Create a customer_complaints table
CREATE TABLE customer_complaints(
  complaint_id int PRIMARY KEY GENERATED BY DEFAULT AS IDENTITY,
  complaint_title text,
  complaint_description text,
  complaint_status text,
  submitted_at timestamp DEFAULT now()
);

Sample Result:

complaint_id	complaint_title	complaint_description	complaint_status	submitted_at
1	Unauthorized transaction	A transaction of $500 was made without my approval.	Pending	2024-09-25 10:15:00
2	Loan denied	My loan was denied despite my good credit score.	Open	2024-09-25 10:16:00

This basic table structure organizes the core details of customer complaints. The complaint_id serves as the primary key, while the complaint_title and complaint_description capture the issue details. The submitted_at timestamp logs when the complaint was filed. This foundation supports adding AI capabilities to enhance how complaints are managed.

Adding AI-Generated Embeddings for Advanced Search

Next, we add an embedding column to store AI-generated vector representations of each complaint. This enables semantic search capabilities, allowing the system to find similar complaints based on meaning rather than keywords.

SQL Query:

-- Add an embeddings column to store AI-generated vector data for complaint_title and complaint_description
ALTER TABLE customer_complaints
    ADD COLUMN complaint_embedding vector(1536) -- Creates vector column with 1536 dimensions
    GENERATED ALWAYS AS  -- Automatically generated on inserts
    (azure_openai.create_embeddings('text-embedding-ada-002', -- Calls Azure OpenAI deployment
    complaint_title || complaint_description)::vector) STORED;  -- Stores the vector

Sample Result:

complaint_id	complaint_title	complaint_description	complaint_status	submitted_at	complaint_embedding
1	Unauthorized transaction	A transaction of $500 was made without my approval.	Pending	2024-09-25 10:15:00	[0.01, -0.03, ...]
2	Loan denied	My loan was denied despite my good credit score.	Open	2024-09-25 10:16:00	[0.08, -0.04, ...]

The complaint_embedding column automatically stores a 1536-dimensional vector, generated by Azure OpenAI, for each complaint. The embeddings are AI-generated vector representations of the complaint details, combining the complaint_title and complaint_description. The 1536-dimensional vectors encode the semantic meaning of the text, enabling the system to understand the context behind each complaint. This AI capability allows for more intelligent searches where similar complaints can be retrieved based on the underlying meaning, even if the wording differs.

Creating a Vector Index for Faster Complaint Search

DiskANN is unique to PostgreSQL on Azure, this index technology stands out as the fastest and most accurate, surpassing industry standards like IVFLAT and HNSW. DiskANN enhances speed by utilizing fast quantized vectors and reduces memory usage by efficiently storing the vector graph on SSD. It maintains high levels of accuracy even as the underlying data grows and evolves, making it a superior solution for large-scale vector search applications.

SQL Query:

-- Create a DiskANN index for fast vector similarity search on the embeddings; 
CREATE INDEX cusomer_complaints_embedding_diskann_idex ON customer complaints USING diskann(embedding vector_cosine_ops);

Inserting Complaints with AI-Generated Embeddings

Let’s insert some sample complaints into the database. As the complaints are added, embeddings will be automatically generated and stored.

SQL Query:

-- Insert sample customer complaints
INSERT INTO customer_complaints (complaint_title, complaint_description, complaint_status)
VALUES
    ('Unauthorized transaction on savings account'
    ,'I noticed a transaction I did not authorize on my savings account.'
    , 'Pending'),
    ('Loan application denied without reason'
    ,'I have been a customer for 10 years, and my loan application was denied without an explanation.'
    , 'Open');

Sample Result:

complaint_id	complaint_title	complaint_description	complaint_status	submitted_at	complaint_embedding
1	Unauthorized transaction on savings account	I noticed a transaction I did not authorize on my savings account.	Pending	2024-09-25 10:30:00	[0.12, -0.02, ...]
2	Loan application denied without reason	I have been a customer for 10 years, and my loan application was denied without explanation.	Open	2024-09-25 10:35:00	[0.09, -0.01, ...]

The embeddings are generated and stored as the rows are inserted, making the complaint data ready for AI-powered searches. As complaints are added to the table, Azure AI automatically generates embeddings based on the text of the complaint title and description. This enables the system to store not just raw data, but also AI-generated vectors that capture the meaning of each complaint. This forms the basis for powerful, AI-driven search capabilities.

Performing a Semantic Search Using Vector Similarity

Now that complaints have been inserted and stored with their vector embeddings, we can perform a semantic search to find the most similar complaints based on a query. For instance, a bank representative may want to search for complaints related to "unauthorized credit card transactions."

SQL Query:

-- Perform a vector similarity search to find the most similar customer complaint to the query

SELECT
 complaint_id, complaint_title, complaint_description
 FROM   customer_complaints c
ORDER BY
    c.complaint_embedding <#> azure_openai.create_embeddings('text-embedding-ada-002', 'Unauthorized credit card transaction')::vector
LIMIT 1;

Sample Result:

complaint_id	complaint_title	complaint_description
1	Unauthorized transaction on savings account	I noticed a transaction I did not authorize on my savings account.

The system performs a semantic search using vector similarity to find the complaint that best matches the query "Unauthorized credit card transaction." This allows the bank to quickly retrieve and handle similar complaints. In this case, the system searches for complaints similar to the phrase "Unauthorized credit card transaction". Although the most relevant complaint in the table mentions a savings account rather than a credit card, the semantic similarity between the two issues (unauthorized transactions) allows the AI model to find the best match. This demonstrates the AI's ability to understand context and meaning rather than relying solely on keywords.

Reference Architecture: Azure Open AI Integration with Azure PostgreSQL Flex

This architecture diagram illustrates a robust banking complaint management system using Azure PostgreSQL Flex integrated with Azure AI.
Users interact with the system via a browser, which communicates through an Azure Application Gateway to an Intelligent App Service.
The app service handles complaint submissions and interacts with the Azure Database for PostgreSQL Flex, where customer complaints are stored as structured data and DiskANN Vector Index Embeddings can be efficiently searched at any scale using the PG_DISCANN extension, which is based on the nearest neighbor search algorithm designed for scalable vector search.
These embeddings enable semantic searches to find similar complaints through SQL queries.
The Azure AI Extension (with services like azure_ai and azure_local_ai) provides AI capabilities, such as vector creation and integration with built-in AI services like Text Analytics, Language Services, and Vision.
Custom AI services, including AI Document Intelligence and Azure Machine Learning, offer advanced analytics, and all data is secured using Disk Encryption Sets.
The architecture ensures data security, AI-powered insights, and efficient management of customer complaints with support for DiskANN Vector to enhance retrieval and search functionalities.

Benefits for the Innovative Applications in Banking Sector

By integrating Azure PostgreSQL Flex and Azure AI, for example, banks can realize several key benefits:

Faster Complaint Resolution: AI-powered semantic search allows bank representatives to find similar complaints quickly and resolve them faster.
Proactive Customer Service: Identifying patterns in customer complaints enables banks to address issues proactively, leading to improved customer satisfaction.
Scalability: Azure PostgreSQL Flex provides the necessary scalability for managing large volumes of customer data.
Advanced Insights: AI models like Azure OpenAI offer deeper insights into customer behavior and complaint trends, helping banks make informed decisions.
Cost Efficiency: Automating the complaint management process reduces operational costs and allows bank representatives to focus on high-value tasks.

Conclusion

Integrating Azure PostgreSQL Flex with Azure AI provides the banking sector with a powerful tool for managing customer complaints. By leveraging AI-generated embeddings and vector similarity searches, banks can streamline the complaint handling process, improve resolution times, and enhance the overall customer experience.

This AI-driven solution not only helps in addressing customer complaints more efficiently but also enables banks to gain valuable insights from recurring issues, ultimately leading to more proactive and personalized customer service.