As organizations more and more undertake AI capabilities throughout their functions, the necessity for centralized administration, safety, and value management of AI mannequin entry is a required step in scaling AI options. The Generative AI Gateway on AWS steering addresses these challenges by offering steering for a unified gateway that helps a number of AI suppliers whereas providing complete governance and monitoring capabilities.
The Generative AI Gateway is a reference structure for enterprises trying to implement end-to-end generative AI options that includes a number of fashions, data-enriched responses, and agent capabilities in a self-hosted approach. This steering combines the broad mannequin entry of Amazon Bedrock, unified developer expertise of Amazon SageMaker AI, and the sturdy administration capabilities of LiteLLM, all whereas supporting buyer entry to fashions from exterior mannequin suppliers in a safer and dependable method.
LiteLLM is an open supply undertaking that addresses frequent challenges confronted by prospects deploying generative AI workloads. LiteLLM simplifies multi-provider mannequin entry whereas standardizing manufacturing operational necessities together with value monitoring, observability, immediate administration, and extra. On this publish we’ll introduce how the Multi-Supplier Generative AI Gateway reference structure supplies steering for deploying LiteLLM into an AWS setting for manufacturing generative AI workload administration and governance.
The problem: Managing multi-provider AI infrastructure
Organizations constructing with generative AI face a number of complicated challenges as they scale their AI initiatives:
- Supplier fragmentation: Groups typically want entry to completely different AI fashions from varied suppliers—Amazon Bedrock, Amazon SageMaker AI, OpenAI, Anthropic, and others—every with completely different APIs, authentication strategies, and billing fashions.
- Decentralized governance mannequin: And not using a unified entry level, organizations battle to implement constant safety insurance policies, utilization monitoring, and value controls throughout completely different AI providers.
- Operational complexity: Managing a number of entry paradigms starting from AWS Identification and Entry Administration roles to API keys, model-specific charge limits, and failover methods throughout suppliers creates operational overhead and will increase the chance of service disruptions.
- Price administration: Understanding and controlling AI spending throughout a number of suppliers and groups turns into more and more tough, notably as utilization scales.
- Safety and compliance: Facilitating constant safety insurance policies and audit trails throughout completely different AI suppliers presents vital challenges for enterprise governance.
Multi-Supplier Generative AI Gateway reference structure
This steering addresses these frequent buyer challenges by offering a centralized gateway that abstracts the complexity of a number of AI suppliers behind a single, managed interface.
Constructed on AWS providers and utilizing the open supply LiteLLM undertaking, organizations can use this answer to combine with AI suppliers whereas sustaining centralized management, safety, and observability.
Versatile deployment choices on AWS
The Multi-Supplier Generative AI Gateway helps a number of deployment patterns to satisfy various organizational wants:
Amazon ECS deployment
For groups preferring containerized functions with managed infrastructure, the ECS deployment supplies serverless container orchestration with computerized scaling and built-in load balancing.
Amazon EKS deployment
Organizations with present Kubernetes experience can use the EKS deployment possibility, which supplies full management over container orchestration whereas benefiting from a managed Kubernetes management aircraft. Clients can deploy a brand new cluster or leverage present clusters for deployment.
The reference structure supplied for these deployment choices is topic to further safety testing primarily based in your group’s particular safety necessities. Conduct further safety testing and assessment as obligatory earlier than deploying something into manufacturing.
Community structure choices
The Multi-Supplier Generative AI Gateway helps a number of community structure choices:
World Public-Dealing with Deployment
For AI providers with international consumer bases, mix the gateway with Amazon CloudFront (CloudFront) and Amazon Route 53. This configuration supplies:
- Enhanced safety with AWS Protect DDoS safety
- Simplified HTTPS administration with the Amazon CloudFront default certificates
- World edge caching for improved latency
- Clever site visitors routing throughout areas
Regional direct entry
For single-Area deployments prioritizing low latency and value optimization, direct entry to the Utility Load Balancer (ALB) removes the CloudFront layer whereas sustaining safety by way of correctly configured safety teams and community ACLs.
Non-public inside entry
Organizations requiring full isolation can deploy the gateway inside a non-public VPC with out web publicity. This configuration makes positive that the AI mannequin entry stays inside your safe community perimeter, with ALB safety teams limiting site visitors to approved personal subnet CIDRs solely.
Complete AI governance and administration
The Multi-Supplier Generative AI Gateway is constructed to allow sturdy AI governance requirements from an easy administrative interface. Along with policy-based configuration and entry administration, customers can configure superior capabilities like load-balancing and immediate caching.
Centralized administration interface
The Generative AI Gateway features a web-based administrative interface in LiteLLM that helps complete administration of LLM utilization throughout your group.
Key capabilities embody:
Person and workforce administration: Configure entry controls at granular ranges, from particular person customers to whole groups, with role-based permissions that align together with your organizational construction.
API key administration: Centrally handle and rotate API keys for the linked AI suppliers whereas sustaining audit trails of key utilization and entry patterns.
Funds controls and alerting: Set spending limits throughout suppliers, groups, and particular person customers with automated alerts when thresholds are approached or exceeded.
Complete value controls: Prices are influenced by AWS infrastructure and LLM suppliers. Whereas it’s the buyer’s duty to configure this answer to satisfy their value necessities, prospects could assessment the present value settings for extra steering.
Helps a number of mannequin suppliers: Suitable with Boto3, OpenAI, and LangGraph SDK, permitting prospects to make use of the most effective mannequin for the workload whatever the supplier.
Help for Amazon Bedrock Guardrails: Clients can leverage guardrails created on Amazon Bedrock Guardrails for his or her generative AI workloads, whatever the mannequin supplier.
Clever routing and resilience
Widespread issues round mannequin deployment embody mannequin and immediate resiliency. These components are necessary to contemplate how failures are dealt with when responding to a immediate or accessing knowledge shops.
Load balancing and failover: The gateway implements refined routing logic that distributes requests throughout a number of mannequin deployments and routinely fails over to backup suppliers when points are detected.
Retry logic: Constructed-in retry mechanisms with exponential back-off facilitate dependable service supply even when particular person suppliers expertise transient points.
Immediate caching: Clever caching helps cut back prices by avoiding duplicate requests to costly AI fashions whereas sustaining response accuracy.
Superior coverage administration
Mannequin deployment structure can vary from the straightforward to extremely complicated. The Multi-Supplier Generative AI Gateway options the superior coverage administration instruments wanted to take care of a robust governance posture.
Fee limiting: Configure refined charge limiting insurance policies that may differ by consumer, API key, mannequin sort, or time of day to facilitate honest useful resource allocation and assist stop abuse.
Mannequin entry controls: Limit entry to particular AI fashions primarily based on consumer roles, ensuring that delicate or costly fashions are solely accessible to approved personnel.
Customized routing guidelines: Implement enterprise logic that routes requests to particular suppliers primarily based on standards similar to request sort, consumer location, or value optimization necessities.
Monitoring and observability
As AI workloads develop to incorporate extra parts, so to do observability wants. The Multi-Supplier Generative AI Gateway structure integrates with Amazon CloudWatch. This integration allows customers to configure myriad monitoring and observability options, together with open-source instruments similar to Langfuse.
Complete logging and analytics
The gateway interactions are routinely logged to CloudWatch, offering detailed insights into:
- Request patterns and utilization developments throughout suppliers and groups
- Efficiency metrics together with latency, error charges, and throughput
- Price allocation and spending patterns by consumer, workforce, and mannequin sort
- Safety occasions and entry patterns for compliance reporting
Constructed-in troubleshooting
The executive interface supplies real-time log viewing capabilities so directors can rapidly diagnose and resolve utilization points while not having to entry CloudWatch immediately.
Amazon SageMaker integration for expanded mannequin entry
Amazon SageMaker helps improve the Multi-Supplier Generative AI Gateway steering by offering a complete machine studying system that seamlessly integrates with the gateway’s structure. Through the use of the Amazon SageMaker managed infrastructure for mannequin coaching, deployment, and internet hosting, organizations can develop customized basis fashions or fine-tune present ones that may be accessed by way of the gateway alongside fashions from different suppliers. This integration removes the necessity for separate infrastructure administration whereas facilitating constant governance throughout each customized and third-party fashions. SageMaker AI mannequin internet hosting capabilities expands the gateway’s mannequin entry to incorporate self-hosted fashions, in addition to these accessible on Amazon Bedrock, OpenAI, and different suppliers.
Our open supply contributions
This reference structure builds upon our contributions to the LiteLLM open supply undertaking, enhancing its capabilities for enterprise deployment on AWS. Our enhancements embody improved error dealing with, enhanced safety features, and optimized efficiency for cloud-native deployments.
Getting began
The Multi-Supplier Generative AI Gateway reference structure is on the market immediately by way of our GitHub repository, full with:
The code repository describes a number of versatile deployment choices to get began.
Public gateway with international CloudFront distribution
Use CloudFront to offer a globally distributed, low-latency entry level on your generative AI providers. The CloudFront edge areas ship content material rapidly to customers around the globe, whereas AWS Protect Customary helps defend towards DDoS assaults. That is the really helpful configuration for public-facing AI providers with a worldwide consumer base.
Customized area with CloudFront
For a extra branded expertise, you possibly can configure the gateway to make use of your personal customized area title, whereas nonetheless benefiting from the efficiency and safety features of CloudFront. This feature is good if you wish to preserve consistency together with your firm’s on-line presence.
Direct entry through public Utility Load Balancer
Clients who prioritize low-latency over international distribution can go for a direct-to-ALB deployment, with out the CloudFront layer. This simplified structure can provide value financial savings, although it requires further consideration for internet utility firewall safety.
Non-public VPC-only entry
For a excessive stage of safety, you possibly can deploy the gateway fully inside a non-public VPC, remoted from the general public web. This configuration is well-suited for processing delicate knowledge or deploying internal-facing generative AI providers. Entry is restricted to trusted networks like VPN, Direct Join, VPC peering, or AWS Transit Gateway.
Study extra and deploy immediately
Able to simplify your multi-provider AI infrastructure? Entry the whole answer bundle to discover an interactive studying expertise with step-by-step steering describing every step of the deployment and administration course of.
Conclusion
The Multi-Supplier Generative AI Gateway is an answer steering supposed to assist prospects get began engaged on generative AI options in a well-architected method, whereas making the most of the AWS setting of providers and complimentary open-source packages. Clients can work with fashions from Amazon Bedrock, Amazon SageMaker JumpStart, or third-party mannequin suppliers. Operations and administration of workloads is carried out through the LiteLLM administration interface, and prospects can select to host on ECS or EKS primarily based on their choice.
As well as, we have now revealed a pattern that integrates the gateway into an agentic customer support utility. The agentic system is orchestrated utilizing LangGraph and deployed on Amazon Bedrock AgentCore. LLM calls are routed by way of the gateway, offering the flexibleness to check brokers with completely different fashions–whether or not hosted on AWS or one other supplier.
This steering is only one a part of a mature generative AI basis on AWS. For deeper studying on the parts of a generative AI system on AWS, see Architect a mature generative AI basis on AWS, which describes further parts of a generative AI system.
Concerning the authors
Dan Ferguson is a Sr. Options Architect at AWS, primarily based in New York, USA. As a machine studying providers knowledgeable, Dan works to help prospects on their journey to integrating ML workflows effectively, successfully, and sustainably.
Bobby Lindsey is a Machine Studying Specialist at Amazon Internet Companies. He’s been in know-how for over a decade, spanning varied applied sciences and a number of roles. He’s presently centered on combining his background in software program engineering, DevOps, and machine studying to assist prospects ship machine studying workflows at scale. In his spare time, he enjoys studying, analysis, mountaineering, biking, and path operating.
Nick McCarthy is a Generative AI Specialist at AWS. He has labored with AWS purchasers throughout varied industries together with healthcare, finance, sports activities, telecoms and power to speed up their enterprise outcomes by way of the usage of AI/ML. Outdoors of labor he likes to spend time touring, making an attempt new cuisines and studying about science and know-how. Nick has a Bachelors diploma in Astrophysics and a Masters diploma in Machine Studying.
Chaitra Mathur is as a GenAI Specialist Options Architect at AWS. She works with prospects throughout industries in constructing scalable generative AI platforms and operationalizing them. All through her profession, she has shared her experience at quite a few conferences and has authored a number of blogs within the Machine Studying and Generative AI domains.
Sreedevi Velagala is a Answer Architect inside the World-Vast Specialist Group Know-how Options workforce at Amazon Internet Companies, primarily based in New Jersey. She has been centered on delivering tailor-made options and steering aligned with the distinctive wants of various clientele throughout AI/ML, Compute, Storage, Networking and Analytics domains. She has been instrumental in serving to prospects learn the way AWS can decrease the compute prices for machine studying workloads utilizing Graviton, Inferentia and Trainium. She leverages her deep technical data and trade experience to ship tailor-made options that align with every shopper’s distinctive enterprise wants and necessities.
