Reference architecture for building a Generative AI (生成AI) workload for financial institutions on AWS, based on the closed-network edition of the AWS Samples "Generative AI Use Cases (GenU)" application with financial-grade security customizations. Use this skill when designing, reviewing, or explaining a financial-services generative AI platform on AWS: RAG-based internal knowledge search, document drafting, customer-support assistance, chat/summarization/translation, image/video generation, closed-network (閉域) deployment, Amazon Bedrock, Amazon Bedrock Guardrails, Amazon Cognito authentication, CloudWatch monitoring dashboards, KMS encryption, and FISC (金融情報システムセンター) compliance mapping for the "BLEA for FSI / 金融リファレンスアーキテクチャ日本版" v1.6 GenAI asset.
Reference architecture for building a Generative AI (生成AI) workload for financial institutions on AWS, based on the closed-network edition of the AWS Samples "Generative AI Use Cases (GenU)" application with financial-grade security customizations. Use this skill when designing, reviewing, or explaining a financial-services generative AI platform on AWS: RAG-based internal knowledge search, document drafting, customer-support assistance, chat/summarization/translation, image/video generation, closed-network (閉域) deployment, Amazon Bedrock, Amazon Bedrock Guardrails, Amazon Cognito authentication, CloudWatch monitoring dashboards, KMS encryption, and FISC (金融情報システムセンター) compliance mapping for the "BLEA for FSI / 金融リファレンスアーキテクチャ日本版" v1.6 GenAI asset.
license
MIT No Attribution
metadata
{"author":"aws-jp-fsi-sa","version":"1.0"}
FSI Reference Architecture — Generative AI (生成AI)
Workload Overview
This skill packages the Generative AI (生成AI) reference architecture from
the "金融リファレンスアーキテクチャ日本版 (BLEA for FSI)" v1.6 initiative. It is a
sample implementation for adopting generative AI inside financial institutions.
The implementation is based on the closed-network (閉域版) edition of the
AWS Samples application Generative AI Use Cases (GenU)
(https://github.com/aws-samples/generative-ai-use-cases), with additional
customizations that address the security requirements financial institutions
typically demand (encryption, audit logging, Guardrail hardening).
GenU is an open-source application that delivers a wide range of generative AI
use cases — chat, text generation, summarization, translation, RAG, image
generation, and video generation. This reference architecture provides:
An organized set of financial-institution use cases achievable with GenU.
An explanation of the main features of the GenU closed-network edition and how
to use them.
The security-hardening customizations applied for financial requirements
(encryption, audit logs, Guardrail settings).
The FISC Safety Standards compliance status (see the FISC compliance summary
below and references/fisc-mapping.md).
Financial use cases addressed include: internal knowledge search & retrieval
via RAG (searching regulations, manuals, past approval documents, and product
materials in natural language), document drafting support (approval documents,
reports, customer-facing materials, internal email), customer-support assistance
(summarizing inquiries, drafting replies, searching similar past cases), data
analysis support (trend analysis of sales/market data, report summarization,
insight extraction), and internal-communication efficiency (automatic meeting
minutes from audio, multi-language translation, material summarization). Image
generation, video generation, and diagram generation extend these to
presentation material, internal training content, and business-flow
visualization.
Prerequisites & positioning. The closed-network configuration is one option
that matches an organization's network policy. The standard (non-closed) edition
of GenU already offers sufficient security features and may be selected depending
on requirements. The design supports staged security hardening: an organization
may start with the standard edition and later add Guardrails, migrate to the
closed-network configuration, or enforce MFA as risk assessment dictates.
Best Practices and Key Components
Best Practices
SHOULD: Confine client-to-system communication to a closed network (閉域) when
data confidentiality or network-security policy requires it; use the GenU
closed-network mode so traffic does not traverse the internet.
SHOULD: Serve the web static frontend via Application Load Balancer + Amazon ECS
Fargate (instead of Amazon CloudFront) when running in the closed-network
configuration.
MUST: Route all AWS Lambda communication to other AWS services (Amazon DynamoDB,
Amazon S3, Amazon Bedrock, etc.) through VPC endpoints so it completes inside the
VPC in the closed-network configuration.
SHOULD: Proxy Amazon Cognito authentication through an Amazon API Gateway Private
API in the closed-network configuration so authentication requests reach the
Cognito User Pool from the isolated environment.
MUST: Authenticate and authorize users with Amazon Cognito; verify JWT token
validity with a Lambda Authorizer so only properly authorized requests reach
backend services.
SHOULD: Consider stricter password policies (minimum length, complexity), MFA
enforcement, and appropriate session-timeout settings on Amazon Cognito
according to organizational requirements.
SHOULD: Provide a CloudWatch monitoring dashboard that visualizes Amazon Bedrock
token usage (input/output tokens), API call status (request count, error rate,
response time), recent prompt history, and user activity patterns.
MUST: Record all operation logs to CloudWatch Logs with appropriate retention
periods and access control for audit and incident investigation.
MUST: Enable Amazon Bedrock Guardrails to filter model input and output; block
sensitive information such as credit card numbers, international bank account
numbers, AWS access keys, email addresses, phone numbers, and passwords.
SHOULD: Customize Guardrails to the organization's policy — e.g., add
Japanese-language support and finance-oriented topic filters (insider trading,
money laundering, regulatory violation, customer confidential information) and
restrict investment-advice or law-violating responses.
WILL: Guardrails are configured independently of the model, so a common policy
can be applied across multiple models.
MUST: Encrypt data at rest (Amazon S3, Amazon DynamoDB, OpenSearch, Guardrails)
with AWS KMS customer-managed keys, and manage KMS key policies and key rotation.
MUST: Enable point-in-time recovery (PITR) on Amazon DynamoDB tables.
PREFER: Serverless managed services (AWS Lambda, Amazon API Gateway, Amazon
DynamoDB, Amazon Bedrock) to reduce operational burden, gain usage-based
auto-scaling, and obtain high availability by default.
WILL: Amazon Bedrock is designed not to use customer data for model training, so
input-data confidentiality is protected.
MUST: Manage the infrastructure as code with AWS CDK so environments are
reproducible across development, verification, and production, and configuration
history is version-controlled for auditability.
MAY: Use the standard (non-closed) edition of GenU instead when full feature
coverage is required, since the closed-network edition does not support external
IdP SAML federation, Voice Chat, or AgentCore Chat, and requires the deploy
region and the model region to be identical.
MUST NOT: Assume the closed-network edition supports SAML federation with an
external IdP, Voice Chat, or AgentCore Chat — these are currently out of scope
for that configuration.
Key Components
Frontend Web Application
Component Overview: Serves the GenU web UI. In the closed-network edition the
static files are distributed by a load balancer and container tasks inside an
isolated subnet instead of a CDN, keeping traffic within the VPC.
Component Overview: Manages user sign-up, sign-in, and request authorization.
In the closed environment a dedicated API Gateway proxy relays authentication
requests to the Cognito User Pool, and a Lambda Authorizer validates JWT tokens.
Component Overview: Hosts the business logic for chat, generation,
summarization, and related use cases, invoked from the frontend and interacting
with data stores and Amazon Bedrock via VPC endpoints.
Assumed AWS Services: Amazon API Gateway, AWS Lambda (VPC-attached).
Generative AI Inference & RAG
Component Overview: Provides foundation-model inference and retrieval-augmented
generation for knowledge search, plus speech and knowledge services used by GenU
use cases. Access from Lambda is through VPC endpoints.
Component Overview: Filters model input and output, blocking sensitive
information and finance-specific denied topics; the hardened configuration adds
Japanese support and finance topic filters.
Assumed AWS Services: Amazon Bedrock Guardrails.
Data Persistence
Component Overview: Stores chat history, system contexts, and generated
artifacts; encrypted with KMS customer-managed keys and protected with PITR on
DynamoDB tables.
Component Overview: Supports speech-to-text and text-to-speech and media
generation features accessible directly from the frontend through VPC endpoints.
Component Overview: Visualizes usage and cost and records operation and model
invocation logs for audit; a dashboard surfaces token usage, invocation counts,
sign-ins, and prompt logs.
This asset is mapped against the FISC Safety Standards (第13版, 13th edition)
practical standards (実務基準). The mapping table records, per standard number,
the controls provided by the common governance-base template, the controls
provided by this GenAI workload template, and additional considerations for the
user. Only measures addressed by BLEA for FSI are listed; overall system safety
must also consider the user application implementation and operational/development
processes.
Representative GenAI-side controls include: user protection via Amazon Cognito
(実1); restriction of client communication to a closed network (実2, 実4, 実34);
IAM-based access control plus AWS KMS encryption for S3/DynamoDB (実3, 実13, 実30);
CloudWatch Logs and VPC Flow Logs (実10, 実72); CDK-managed configuration with
DynamoDB PITR enabled (実39, 実42); Amazon Bedrock Guardrails for content
filtering (実69, 実152); and CloudWatch dashboard monitoring plus CloudTrail
auditing for AI-usage governance (実151, 実152). Several standards are marked out
of scope because they concern financial-transaction services, management/operation
processes, development processes, or are already handled on the AWS cloud side.
The complete, unabridged table (every 実N entry) is preserved in
references/fisc-mapping.md.
Reference Architecture
The closed-network architecture runs entirely inside an Amazon VPC. A browser in
an isolated subnet reaches the frontend (Application Load Balancer + Amazon ECS
Fargate); AWS Lambda functions and the frontend reach AWS services (Amazon
DynamoDB and Amazon S3 via Gateway VPC endpoints; Amazon Bedrock, Amazon
Transcribe, Amazon Polly, Amazon Cognito, Amazon Kendra, and others via Interface
VPC endpoints); authentication is proxied through an Amazon API Gateway Private
API to Amazon Cognito; and a Route 53 Resolver endpoint provides in-VPC name
resolution.
See assets/images/arch.png for the diagram and references/architecture.md for
the full technical detail (features, security hardening, architectural
characteristics, and closed-network feature limitations).
CDK Sample Overview
No CDK sample is provided for this asset in BLEA for FSI v1.6; this asset is
documentation only.
The BLEA for FSI repository does not contain a CDK application for this workload.
Instead, the sample is realized by cloning the external AWS Samples GenU
repository at a specific commit and applying a customization patch
(changes.diff) shipped in the source documentation. The deployment procedure
(clone/patch/build/deploy of the external GenU project, Bedrock model-invocation
logging setup, and verification) is captured, for reference, in
references/deployment.md. The actual CDK code being deployed lives in the
external GenU repository, not in this asset.
