Enterprises are searching for to rapidly unlock the potential of generative AI by offering entry to basis fashions (FMs) to totally different strains of enterprise (LOBs). IT groups are accountable for serving to the LOB innovate with pace and agility whereas offering centralized governance and observability. For instance, they could want to trace the utilization of FMs throughout groups, chargeback prices and supply visibility to the related price middle within the LOB. Moreover, they could want to manage entry to totally different fashions per staff. For instance, if solely particular FMs could also be accepted to be used.
Amazon Bedrock is a completely managed service that provides a selection of high-performing basis fashions from main AI corporations like AI21 Labs, Anthropic, Cohere, Meta, Stability AI, and Amazon by way of a single API, together with a broad set of capabilities to construct generative AI functions with safety, privateness, and accountable AI. As a result of Amazon Bedrock is serverless, you don’t must handle any infrastructure, and you may securely combine and deploy generative AI capabilities into your functions utilizing the AWS providers you might be already conversant in.
A software program as a service (SaaS) layer for basis fashions can present a easy and constant interface for end-users, whereas sustaining centralized governance of entry and consumption. API gateways can present unfastened coupling between mannequin shoppers and the mannequin endpoint service, and suppleness to adapt to altering mannequin, architectures, and invocation strategies.
On this publish, we present you how one can construct an inner SaaS layer to entry basis fashions with Amazon Bedrock in a multi-tenant (staff) structure. We particularly concentrate on utilization and value monitoring per tenant and likewise controls reminiscent of utilization throttling per tenant. We describe how the answer and Amazon Bedrock consumption plans map to the overall SaaS journey framework. The code for the answer and an AWS Cloud Growth Equipment (AWS CDK) template is obtainable within the GitHub repository.
Challenges
An AI platform administrator wants to offer standardized and quick access to FMs to a number of improvement groups.
The next are among the challenges to offer ruled entry to basis fashions:
- Value and utilization monitoring – Monitor and audit particular person tenant prices and utilization of basis fashions, and supply chargeback prices to particular price facilities
- Price range and utilization controls – Handle API quota, finances, and utilization limits for the permitted use of basis fashions over an outlined frequency per tenant
- Entry management and mannequin governance – Outline entry controls for particular permit listed fashions per tenant
- Multi-tenant standardized API – Present constant entry to basis fashions with OpenAPI requirements
- Centralized administration of API – Present a single layer to handle API keys for accessing fashions
- Mannequin variations and updates – Deal with new and up to date mannequin model rollouts
Answer overview
On this resolution, we confer with a multi-tenant strategy. A tenant right here can vary from a person consumer, a selected challenge, staff, and even a whole division. As we talk about the strategy, we use the time period staff, as a result of it’s the commonest. We use API keys to limit and monitor API entry for groups. Every staff is assigned an API key for entry to the FMs. There might be totally different consumer authentication and authorization mechanisms deployed in a corporation. For simplicity, we don’t embrace these on this resolution. You might also combine current identification suppliers with this resolution.
The next diagram summarizes the answer structure and key elements. Groups (tenants) assigned to separate price facilities eat Amazon Bedrock FMs by way of an API service. To trace consumption and value per staff, the answer logs knowledge for every particular person invocation, together with the mannequin invoked, variety of tokens for textual content era fashions, and picture dimensions for multi-modal fashions. As well as, it aggregates the invocations per mannequin and prices by every staff.
You’ll be able to deploy the answer in your personal account utilizing the AWS CDK. AWS CDK is an open supply software program improvement framework to mannequin and provision your cloud utility sources utilizing acquainted programming languages. The AWS CDK code is obtainable within the GitHub repository.
Within the following sections, we talk about the important thing elements of the answer in additional element.
Capturing basis mannequin utilization per staff
The workflow to seize FM utilization per staff consists of the next steps (as numbered within the previous diagram):
- A staff’s utility sends a POST request to Amazon API Gateway with the mannequin to be invoked within the
model_id
question parameter and the consumer immediate within the request physique. - API Gateway routes the request to an AWS Lambda operate (
bedrock_invoke_model
) that’s accountable for logging staff utilization info in Amazon CloudWatch and invoking the Amazon Bedrock mannequin. - Amazon Bedrock gives a VPC endpoint powered by AWS PrivateLink. On this resolution, the Lambda operate sends the request to Amazon Bedrock utilizing PrivateLink to determine a non-public connection between the VPC in your account and the Amazon Bedrock service account. To be taught extra about PrivateLink, see Use AWS PrivateLink to arrange non-public entry to Amazon Bedrock.
- After the Amazon Bedrock invocation, Amazon CloudTrail generates a CloudTrail occasion.
- If the Amazon Bedrock name is profitable, the Lambda operate logs the next info relying on the kind of invoked mannequin and returns the generated response to the appliance:
- team_id – The distinctive identifier for the staff issuing the request.
- requestId – The distinctive identifier of the request.
- model_id – The ID of the mannequin to be invoked.
- inputTokens – The variety of tokens despatched to the mannequin as a part of the immediate (for textual content era and embeddings fashions).
- outputTokens – The utmost variety of tokens to be generated by the mannequin (for textual content era fashions).
- peak – The peak of the requested picture (for multi-modal fashions and multi-modal embeddings fashions).
- width – The width of the requested picture (for multi-modal fashions solely).
- steps – The steps requested (for Stability AI fashions).
Monitoring prices per staff
A distinct circulate aggregates the utilization info, then calculates and saves the on-demand prices per staff every day. By having a separate circulate, we be certain that price monitoring doesn’t impression the latency and throughput of the mannequin invocation circulate. The workflow steps are as follows:
- An Amazon EventBridge rule triggers a Lambda operate (
bedrock_cost_tracking
) day by day. - The Lambda operate will get the utilization info from CloudWatch for yesterday, calculates the related prices, and shops the info aggregated by
team_id
andmodel_id
in Amazon Easy Storage Service (Amazon S3) in CSV format.
To question and visualize the info saved in Amazon S3, you’ve gotten totally different choices, together with S3 Choose, and Amazon Athena and Amazon QuickSight.
Controlling utilization per staff
A utilization plan specifies who can entry a number of deployed APIs and optionally units the goal request charge to begin throttling requests. The plan makes use of API keys to establish API purchasers who can entry the related API for every key. You need to use API Gateway utilization plans to throttle requests that exceed predefined thresholds. You may also use API keys and quota limits, which allow you to set the utmost variety of requests per API key every staff is permitted to challenge inside a specified time interval. That is along with Amazon Bedrock service quotas which can be assigned solely on the account stage.
Conditions
Earlier than you deploy the answer, be sure to have the next:
Deploy the AWS CDK stack
Comply with the directions within the README file of the GitHub repository to configure and deploy the AWS CDK stack.
The stack deploys the next sources:
- Personal networking setting (VPC, non-public subnets, safety group)
- IAM position for controlling mannequin entry
- Lambda layers for the mandatory Python modules
- Lambda operate
invoke_model
- Lambda operate
list_foundation_models
- Lambda operate
cost_tracking
- Relaxation API (API Gateway)
- API Gateway utilization plan
- API key related to the utilization plan
Onboard a brand new staff
For offering entry to new groups, you’ll be able to both share the identical API key throughout totally different groups and monitor the mannequin consumptions by offering a unique team_id
for the API invocation, or create devoted API keys used for accessing Amazon Bedrock sources by following the directions supplied within the README.
The stack deploys the next sources:
- API Gateway utilization plan related to the beforehand created REST API
- API key related to the utilization plan for the brand new staff, with reserved throttling and burst configurations for the API
For extra details about API Gateway throttling and burst configurations, confer with Throttle API requests for higher throughput.
After you deploy the stack, you’ll be able to see that the brand new API key for team-2
is created as nicely.
Configure mannequin entry management
The platform administrator can permit entry to particular basis fashions by enhancing the IAM coverage related to the Lambda operate invoke_model
. The
IAM permissions are outlined within the file setup/stack_constructs/iam.py. See the next code:
Invoke the service
After you’ve gotten deployed the answer, you’ll be able to invoke the service immediately out of your code. The next
is an instance in Python for consuming the invoke_model
API for textual content era via a POST request:
Output: Amazon Bedrock is an inner expertise platform developed by Amazon to run and function a lot of their providers and merchandise. Some key issues about Bedrock …
The next is one other instance in Python for consuming the invoke_model
API for embeddings era via a POST request:
model_id = "amazon.titan-embed-text-v1" #the mannequin id for the Amazon Titan Embeddings Textual content mannequin
immediate = "What's Amazon Bedrock?"
response = requests.publish(
f"{api_url}/invoke_model?model_id={model_id}",
json={"inputs": immediate, "parameters": model_kwargs},
headers={
"x-api-key": api_key, #key for querying the API
"team_id": team_id #distinctive tenant identifier,
"embeddings": "true" #boolean worth for the embeddings mannequin
}
)
textual content = response.json()[0]["embedding"]
Output: 0.91796875, 0.45117188, 0.52734375, -0.18652344, 0.06982422, 0.65234375, -0.13085938, 0.056884766, 0.092285156, 0.06982422, 1.03125, 0.8515625, 0.16308594, 0.079589844, -0.033935547, 0.796875, -0.15429688, -0.29882812, -0.25585938, 0.45703125, 0.044921875, 0.34570312 …
Entry denied to basis fashions
The next is an instance in Python for consuming the invoke_model
API for textual content era via a POST request with an entry denied response:
<Response [500]> “Traceback (most up-to-date name final):n File ”/var/job/index.py”, line 213, in lambda_handlern response = _invoke_text(bedrock_client, model_id, physique, model_kwargs)n File ”/var/job/index.py”, line 146, in _invoke_textn increase en File ”/var/job/index.py”, line 131, in _invoke_textn response = bedrock_client.invoke_model(n File ”/choose/python/botocore/consumer.py”, line 535, in _api_calln return self._make_api_call(operation_name, kwargs)n File ”/choose/python/botocore/consumer.py”, line 980, in _make_api_calln increase error_class(parsed_response, operation_name)nbotocore.errorfactory.AccessDeniedException: An error occurred (AccessDeniedException) when calling the InvokeModel operation: Your account is just not licensed to invoke this API operation.n”
Value estimation instance
When invoking Amazon Bedrock fashions with on-demand pricing, the entire price is calculated because the sum of the enter and output prices. Enter prices are based mostly on the variety of enter tokens despatched to the mannequin, and output prices are based mostly on the tokens generated. The costs are per 1,000 enter tokens and per 1,000 output tokens. For extra particulars and particular mannequin costs, confer with Amazon Bedrock Pricing.
Let’s take a look at an instance the place two groups, team1 and team2, entry Amazon Bedrock via the answer on this publish. The utilization and value knowledge saved in Amazon S3 in a single day is proven within the following desk.
The columns input_tokens
and output_tokens
retailer the entire enter and output tokens throughout mannequin invocations per mannequin and per staff, respectively, for a given day.
The columns input_cost
and output_cost
retailer the respective prices per mannequin and per staff. These are calculated utilizing the next formulation:
input_cost = input_token_count * model_pricing["input_cost"] / 1000
output_cost = output_token_count * model_pricing["output_cost"] / 1000
team_id | model_id | input_tokens | output_tokens | invocations | input_cost | output_cost |
Team1 | amazon.titan-tg1-large | 24000 | 2473 | 1000 | 0.0072 | 0.00099 |
Team1 | anthropic.claude-v2 | 2448 | 4800 | 24 | 0.02698 | 0.15686 |
Team2 | amazon.titan-tg1-large | 35000 | 52500 | 350 | 0.0105 | 0.021 |
Team2 | ai21.j2-grande-instruct | 4590 | 9000 | 45 | 0.05738 | 0.1125 |
Team2 | anthropic.claude-v2 | 1080 | 4400 | 20 | 0.0119 | 0.14379 |
Finish-to-end view of a purposeful multi-tenant serverless SaaS setting
Let’s perceive what an end-to-end purposeful multi-tenant serverless SaaS setting would possibly seem like. The next is a reference structure diagram.
This structure diagram is a zoomed-out model of the earlier structure diagram defined earlier within the publish, the place the earlier structure diagram explains the small print of one of many microservices talked about (foundational mannequin service). This diagram explains that, other than foundational mannequin service, you might want to produce other elements as nicely in your multi-tenant SaaS platform to implement a purposeful and scalable platform.
Let’s undergo the small print of the structure.
Tenant functions
The tenant functions are the entrance finish functions that work together with the setting. Right here, we present a number of tenants accessing from totally different native or AWS environments. The entrance finish functions might be prolonged to incorporate a registration web page for brand new tenants to register themselves and an admin console for directors of the SaaS service layer. If the tenant functions require a customized logic to be carried out that wants interplay with the SaaS setting, they’ll implement the specs of the appliance adaptor microservice. Instance eventualities might be including customized authorization logic whereas respecting the authorization specs of the SaaS setting.
Shared providers
The next are shared providers:
- Tenant and consumer administration providers –These providers are accountable for registering and managing the tenants. They supply the cross-cutting performance that’s separate from utility providers and shared throughout all the tenants.
- Basis mannequin service –The answer structure diagram defined originally of this publish represents this microservice, the place the interplay from API Gateway to Lambda capabilities is going on throughout the scope of this microservice. All tenants use this microservice to invoke the foundations fashions from Anthropic, AI21, Cohere, Stability, Meta, and Amazon, in addition to fine-tuned fashions. It additionally captures the data wanted for utilization monitoring in CloudWatch logs.
- Value monitoring service –This service tracks the associated fee and utilization for every tenant. This microservice runs on a schedule to question the CloudWatch logs and output the aggregated utilization monitoring and inferred price to the info storage. The price monitoring service might be prolonged to construct additional studies and visualization.
Utility adaptor service
This service presents a set of specs and APIs {that a} tenant could implement as a way to combine their customized logic to the SaaS setting. Based mostly on how a lot customized integration is required, this element might be non-compulsory for tenants.
Multi-tenant knowledge retailer
The shared providers retailer their knowledge in a knowledge retailer that may be a single shared Amazon DynamoDB desk with a tenant partitioning key that associates DynamoDB gadgets with particular person tenants. The price monitoring shared service outputs the aggregated utilization and value monitoring knowledge to Amazon S3. Based mostly on the use case, there might be an application-specific knowledge retailer as nicely.
A multi-tenant SaaS setting can have much more elements. For extra info, confer with Constructing a Multi-Tenant SaaS Answer Utilizing AWS Serverless Providers.
Help for a number of deployment fashions
SaaS frameworks sometimes define two deployment fashions: pool and silo. For the pool mannequin, all tenants entry FMs from a shared setting with widespread storage and compute infrastructure. Within the silo mannequin, every tenant has its personal set of devoted sources. You’ll be able to examine isolation fashions within the SaaS Tenant Isolation Methods whitepaper.
The proposed resolution might be adopted for each SaaS deployment fashions. Within the pool strategy, a centralized AWS setting hosts the API, storage, and compute sources. In silo mode, every staff accesses APIs, storage, and compute sources in a devoted AWS setting.
The answer additionally suits with the out there consumption plans supplied by Amazon Bedrock. AWS gives a selection of two consumptions plan for inference:
- On-Demand – This mode permits you to use basis fashions on a pay-as-you-go foundation with out having to make any time-based time period commitments
- Provisioned Throughput – This mode permits you to provision enough throughput to fulfill your utility’s efficiency necessities in alternate for a time-based time period dedication
For extra details about these choices, confer with Amazon Bedrock Pricing.
The serverless SaaS reference resolution described on this publish can apply the Amazon Bedrock consumption plans to offer primary and premium tiering choices to end-users. Fundamental might embrace On-Demand or Provisioned Throughput consumption of Amazon Bedrock and will embrace particular utilization and finances limits. Tenant limits might be enabled by throttling requests based mostly on requests, token sizes, or finances allocation. Premium tier tenants might have their very own devoted sources with provisioned throughput consumption of Amazon Bedrock. These tenants would sometimes be related to manufacturing workloads that require excessive throughput and low latency entry to Amazon Bedrock FMs.
Conclusion
On this publish, we mentioned how one can construct an inner SaaS platform to entry basis fashions with Amazon Bedrock in a multi-tenant setup with a concentrate on monitoring prices and utilization, and throttling limits for every tenant. Extra matters to discover embrace integrating current authentication and authorization options within the group, enhancing the API layer to incorporate net sockets for bi-directional consumer server interactions, including content material filtering and different governance guardrails, designing a number of deployment tiers, integrating different microservices within the SaaS structure, and lots of extra.
Your entire code for this resolution is obtainable within the GitHub repository.
For extra details about SaaS-based frameworks, confer with SaaS Journey Framework: Constructing a New SaaS Answer on AWS.
In regards to the Authors
Hasan Poonawala is a Senior AI/ML Specialist Options Architect at AWS, working with Healthcare and Life Sciences clients. Hasan helps design, deploy and scale Generative AI and Machine studying functions on AWS. He has over 15 years of mixed work expertise in machine studying, software program improvement and knowledge science on the cloud. In his spare time, Hasan likes to discover nature and spend time with family and friends.
Anastasia Tzeveleka is a Senior AI/ML Specialist Options Architect at AWS. As a part of her work, she helps clients throughout EMEA construct basis fashions and create scalable generative AI and machine studying options utilizing AWS providers.
Bruno Pistone is a Generative AI and ML Specialist Options Architect for AWS based mostly in Milan. He works with massive clients serving to them to deeply perceive their technical wants and design AI and Machine Studying options that make the very best use of the AWS Cloud and the Amazon Machine Studying stack. His experience embrace: Machine Studying finish to finish, Machine Studying Industrialization, and Generative AI. He enjoys spending time along with his mates and exploring new locations, in addition to travelling to new locations.
Vikesh Pandey is a Generative AI/ML Options architect, specialising in monetary providers the place he helps monetary clients construct and scale Generative AI/ML platforms and resolution which scales to a whole lot to even 1000’s of customers. In his spare time, Vikesh likes to jot down on varied weblog boards and construct legos along with his child.