Project: AI Personal Financial Analyst

Project Description

This project implements a personal financial analyst system that leverages a large language model (LLM), specifically based on GPT-4o and DeepSeek R1, to make informed decisions in the financial market. The system collects real-time financial data, such as stock and cryptocurrency information, stores this data in a MongoDB database, and subsequently converts the data into embeddings that are stored in a database with PGVector support. The LLM is then fed with these embeddings to provide accurate and personalized financial insights.

Project Structure

The project is divided into several main components, each responsible for a specific functionality:

• Data Collection: Utilizes an external client (AlphaClient) to fetch financial data for stocks and cryptocurrencies.
• Data Storage: The collected data is stored in MongoDB for persistence.
• Embedding Conversion: The data is converted into embeddings using the LLM and stored in a PGVector-supported database.
• LLM Integration: The LLM (based on GPT-4 and DeepSeek R1) uses these embeddings to generate financial insights and responses.

Key Components

• DataLoadingService: A service responsible for loading stock and cryptocurrency data, processing it, and storing it both in MongoDB and PGVector. This service uses the StockEnum and CryptoEnum enums to iterate over the entities that need to be processed.

• AlphaClient: A client responsible for communicating with an external API to fetch financial information for stocks and cryptocurrencies.

• Stock and Crypto: Model classes representing the stock and cryptocurrency entities, respectively. They store the necessary data that will be used by the LLM.

• LLMContent: A component that prepares the content to be used by the LLM to generate embeddings.

• AIFinancialRepository: The AIFinancialRepository class integrates embedding generation, vector storage, and retrieval augmentation to index documents with metadata and retrieve similar documents based on user queries. It leverages a chat model, query transformer, document retriever, and query augmenter to support a Retrieval Augmented Generation (RAG) approach. This enables context-rich, similarity-based search in financial applications.

Dependencies

The project uses several libraries and frameworks to ensure its functionality. Below are the main dependencies and their versions as specified in the pom.xml:

• Spring Boot: The main framework used to create the service. It provides easy integration with MongoDB, JPA, and other RESTful services.

  • Version: 3.4.2

• MongoDB: Used to store the collected financial data.

  • Driver: spring-boot-starter-data-mongodb
  • Version: 3.4.2

• PGVector: Used to store the embeddings generated from the financial data.

  • Dependency: spring-ai-pgvector-store (requires specific configuration in PostgreSQL to support vectors)
  • Version: 1.0.0-M5

• SLF4J + Logback: Used for logging within the project.

  • Version: 2.23.1

• JUnit & Mockito: Testing tools used to create unit and integration tests.

  • JUnit Version: 5.10.3
  • Mockito Version: 5.11.0

• Spring AI: Spring AI OpenAI provides a streamlined approach to developing AI with gpt model to powered applications using the Spring framework.

  • Dependency: spring-ai-openai-spring-boot-starter
  • Version: 1.0.0-M5

• Spring AI: Spring AI Ollama provides a streamlined approach to developing AI with ollama and opensource models. Implementation ready to support differents LLM models: DeepSeek, Mistral, Gemma, Ollama, etc.

  • Dependency: spring-ai-ollama-spring-boot-starter
  • Version: 1.0.0-M5

Configuration and Execution

Prerequisites

1. Java 22: The project uses features of Java 22, so ensure you have this version or higher installed.
2. MongoDB: Required to store financial data. Must be configured and running.
3. PostgreSQL with PGVector: Must have the PGVector plugin installed and configured in PostgreSQL.
4. OpenAI API Key: If you want use OpenAI model, Required to access the GPT-4o model. Configure your API key in the application properties.
5. Alpha API Key: Required to access the Financial Stock Market API, providing developers with access to a diverse range of financial data.

Execution Steps

1. Clone the repository.
2. Configure the properties in the application.properties file to point to your MongoDB and PostgreSQL instances.
3. Run docker compose to start local PGVector and MongoDB:

   docker-compose up -d

4. After start docker-compose you can access Mongo Express in this url: http://localhost:8081
5. Build the project using Maven:

   mvn clean install

6. Run the project:

   mvn spring-boot:run

Config Keycloack https://medium.com/javarevisited/keycloak-integration-with-spring-security-6-37999f43ec85

Execute application local

mvn spring-boot:run

Request Curls

Generate token

curl --location 'http://localhost:8080/realms/ai-financial-analyst/protocol/openid-connect/token' \
--header 'Content-Type: application/x-www-form-urlencoded' \
--data-urlencode 'grant_type=password' \
--data-urlencode 'client_id=super-client' \
--data-urlencode 'username=user1' \
--data-urlencode 'password=123456'

Load Data

Load Data in Vector DB and MongoDB. Just need run when you want update database. If you use Postman, you will need get access token on generate token response and inject this token like Bearer Token

curl --location 'http://localhost:9090/load' \
--header 'Authorization: ••••••'

Chat

Request to talk with chatbot, you will need do the same step of security. Inject token on request.

curl --location 'http://localhost:9090/chat/question' \
--header 'Content-Type: application/json' \
--header 'Authorization: Bearer eyJhbGciOiJSUzI1NiIsInR5cCIgOiAiSldUIiwia2lkIiA6ICJ3OGV1eTNJRUxJZTZ5c2ZvYnJKaGF6NF9lRndVWXNCSVpDcDVPVUs5MkFnIn0.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.X0hVXCXYgWSUkPlQmZo5Jam1QyJ-AD96BO9g7z7bexHOOflB2i6HLKZOyglM-9kIkdMJeABKEo-a3tH14TM8Air4W8WBYiaAT1ar2JmaODtBN1Q5KGRfdWV-Z7-WcHAQ0R4SB62BviocmahS6bwl3eOa9U5EYQtIWnBLVEIxMQRzCTwPgBhPMeUD1GLcMvRy53OXEZqgL5TC8s6Pc5c80lho__KOsh9KtalbLHLvZA-HzN3ZPJt9SUpogUPl5_W8c4qd-5eJBllHfTLbD_PMHADoutI-pJ1P4ZH0Weo1yYDiPQW8pShbUoLvCUgDeXCslRxIji9TZgKcCxWisP22MQ' \
--data '{
    "question": " What Are All 10 Stocks Warren Buffett Is Selling? Traduza para portugues a resposta"
}'

AI Financial Analyst Design System

Next Steps

1. Benchmarks between reasoning models vs general models: Explore enhancements in preparing the data for the LLM, including adding more contextual data to improve the generated responses.
2. Discover new concepts of RAGStorage Optimization: Best improvent to prevent hallucinations.
3. Serverless Architecute: Check cost and performance using different cloud providers like: AWS, Azure or GCP.

Medium Story

GenAI Design Patterns to Build a Next-Generation Financial Analysis Application and IA Benchmark