Amazon Bedrock & Foundation Models in Yemen

The Amazon Bedrock & Foundation Models course enables learners to design an AI-powered banking assistant that helps customers and banking employees access banking information, understand financial products, and find answers to common service-related questions. Using Amazon Bedrock, foundation models, prompt engineering, and Retrieval-Augmented Generation (RAG), learners can develop a solution that retrieves relevant information from approved banking documents and generates context-aware responses.

2. Problem Statement

Banks manage large volumes of information, including account service guidelines, loan policies, credit card terms, interest rate documents, frequently asked questions, and internal operational procedures. Customers often experience delays when searching for answers, while banking employees spend significant time locating information across multiple documents.

Traditional rule-based chatbots may struggle with complex questions, varied wording, and requests requiring information from several documents. Incorrect or unsupported responses can also create compliance and customer-trust risks.

Key challenges identified:

• Difficulty accessing accurate banking information quickly.
• High volume of repetitive customer service queries.
• Limited ability of traditional chatbots to understand natural-language questions.
• Time-consuming searches across banking policies and product documents.
• Risk of generating inaccurate or unsupported financial information.
• Need to protect confidential customer data and enforce access controls.

3. How Did We Overcome the Challenges?

Learners applied Amazon Bedrock and foundation model capabilities to develop a controlled, document-grounded banking knowledge assistant.

• Foundation Model Integration: Used an appropriate model available through Amazon Bedrock to understand questions and generate clear, relevant responses.
• Prompt Engineering: Designed prompts that instruct the assistant to answer using approved banking information, avoid unsupported claims, and request clarification when necessary.
• Retrieval-Augmented Generation (RAG): Retrieved relevant passages from approved banking documents before generating answers, helping ground responses in the bank's source material.
• Document Processing: Organised banking FAQs, loan guidelines, credit card information, and service procedures into searchable content, with suitable chunking and metadata.
• Knowledge Retrieval: Used embeddings and semantic search to find information relevant to a customer's question, including questions phrased differently from the source documents.
• Security and Access Control: Applied appropriate AWS Identity and Access Management (IAM) permissions and data-handling controls. Restricted access to confidential information and excluded sensitive customer records from the demonstration dataset.
• Response Validation: Tested answers against source documents, checked relevance and consistency, and configured the assistant to acknowledge when information was unavailable.
• Monitoring and Cost Awareness: Reviewed response latency, model usage, and token consumption to support responsible and cost-conscious implementation.

4. Overall Implementation

Step 1: Define the Banking Use Case

Selected a banking support assistant focused on frequently asked questions about savings accounts, loan eligibility guidelines, credit cards, service procedures, and general banking policies. Personalised financial advice, transaction execution, and account-specific disclosures were excluded from the initial prototype.

Step 2: Prepare Banking Documents

Collected sample, non-confidential banking documents such as product brochures, service FAQs, loan policy summaries, and customer support guidelines. Cleaned and organised the documents for retrieval.

Step 3: Configure Amazon Bedrock

Set up the required AWS environment, reviewed model access and permissions, selected a suitable foundation model, and configured the necessary SDK or API integration.

Step 4: Implement Document Retrieval

Processed documents into searchable chunks and generated embeddings using a compatible embedding model. Stored the embeddings in an appropriate vector store or configured Amazon Bedrock Knowledge Bases where suitable and available.

Step 5: Develop the AI Assistant

Built an application that accepts a natural-language question, retrieves relevant banking information, and supplies that context to the foundation model. The model then generates an answer based on the retrieved information and the defined prompt instructions.

Step 6: Apply Security and Responsible AI Controls

Configured least-privilege access, protected credentials, minimised sensitive data, and added instructions to avoid unsupported financial claims. Designed the assistant to refer users to authorised banking staff for account-specific, sensitive, or high-risk requests.

Step 7: Test and Evaluate the Solution

Tested common questions, differently worded questions, ambiguous requests, missing information, and questions outside the approved document collection. Compared generated responses with source material and recorded issues for improvement.

Step 8: Demonstrate the Capstone Project

Presented a working prototype that demonstrates document ingestion, semantic retrieval, foundation model integration, response generation, error handling, and basic evaluation. Documented the architecture, test cases, observed limitations, and recommendations for further development.

5. Evidence of Learner Outcomes

Learners can demonstrate their skills through the following project artefacts:

• Problem Analysis: Documented banking support challenges, user needs, and project scope.
• AWS Configuration: Evidence of the configured development environment, permissions, and model invocation workflow, with credentials and sensitive details removed.
• Document Preparation: Sample document-processing pipeline and examples of searchable document chunks.
• RAG Implementation: Demonstration of retrieving relevant passages and using them to generate grounded responses.
• Prompt Engineering: Prompt templates and test examples showing how instructions affect response quality.
• API Integration: Code or execution logs demonstrating successful requests to Amazon Bedrock, with secrets and personal information redacted.
• Response Evaluation: Test cases comparing generated answers with approved source documents, including examples of correct answers and identified limitations.
• Security Awareness: Documentation of access restrictions, data-minimisation measures, and handling of unsupported or sensitive requests.
• Capstone Demonstration: A working banking knowledge assistant with example questions, retrieved evidence, generated answers, and documented results.

6. Expected Learning Outcomes

After completing the project, learners should be able to:

• Explain foundation models and their role in Generative AI applications.
• Integrate an available foundation model using Amazon Bedrock APIs or SDKs.
• Apply prompt engineering techniques to banking information queries.
• Build a basic RAG workflow using document processing, embeddings, and semantic retrieval.
• Connect retrieved banking information with model-generated responses.
• Evaluate answer relevance, grounding, consistency, and limitations.
• Apply foundational cloud security and responsible AI practices.
• Identify cost, latency, privacy, and governance considerations for banking AI applications.
• Present a documented AI project demonstrating practical application of Amazon Bedrock and foundation models.

7. Final Project Outcome

The final deliverable is a prototype AI-Powered Banking Knowledge Assistant that helps users find answers to general banking questions using approved source documents. The project demonstrates how Amazon Bedrock, foundation models, prompt engineering, embeddings, and RAG can be combined to improve information discovery and support service workflows.

The prototype is intended for learning and demonstration. Before production use, a bank would need formal security and compliance reviews, approved data access, rigorous model evaluation, monitoring, auditability, and human escalation procedures. It should not be treated as a system for executing financial transactions, making credit decisions, or providing personalised financial advice without the appropriate controls and approvals.