Use Amazon SageMaker Unified Studio to construct complicated AI workflows utilizing Amazon Bedrock Flows

Organizations face the problem to handle information, a number of synthetic intelligence and machine studying (AI/ML) instruments, and workflows throughout completely different environments, impacting productiveness and governance. A unified improvement surroundings consolidates information processing, mannequin improvement, and AI utility deployment right into a single system. This integration streamlines workflows, enhances collaboration, and accelerates AI resolution improvement from idea to manufacturing.

The subsequent technology of Amazon SageMaker is the middle on your information, analytics, and AI. SageMaker brings collectively AWS AI/ML and analytics capabilities and delivers an built-in expertise for analytics and AI with unified entry to information. Amazon SageMaker Unified Studio is a single information and AI improvement surroundings the place you will discover and entry your information and act on it utilizing AWS analytics and AI/ML companies, for SQL analytics, information processing, mannequin improvement, and generative AI utility improvement.

With SageMaker Unified Studio, you’ll be able to effectively construct generative AI functions in a trusted and safe surroundings utilizing Amazon Bedrock. You may select from a number of high-performing basis fashions (FMs) and superior customization and tooling equivalent to Amazon Bedrock Information Bases, Amazon Bedrock Guardrails, Amazon Bedrock Brokers, and Amazon Bedrock Flows. You may quickly tailor and deploy generative AI functions, and share with the built-in catalog for discovery.

On this submit, we exhibit how you should utilize SageMaker Unified Studio to create complicated AI workflows utilizing Amazon Bedrock Flows.

Answer overview

Take into account FinAssist Corp, a number one monetary establishment growing a generative AI-powered agent help utility. The answer affords the next key options:

• Criticism reference system – An AI-powered system offering fast entry to historic grievance information, enabling customer support representatives to effectively deal with buyer follow-ups, help inside audits, and support in coaching new employees.
• Clever data base – A complete information supply of resolved complaints that rapidly retrieves related grievance particulars, decision actions, and consequence summaries.
• Streamlined workflow administration – Enhanced consistency in buyer communications by standardized entry to previous case data, supporting compliance checks and course of enchancment initiatives.
• Versatile question functionality – A simple interface supporting varied question situations, from buyer inquiries about previous resolutions to inside critiques of grievance dealing with procedures.

Let’s discover how SageMaker Unified Studio and Amazon Bedrock Flows, built-in with Amazon Bedrock Information Bases and Amazon Bedrock Brokers, tackle these challenges by creating an AI-powered grievance reference system. The next diagram illustrates the answer structure.

The answer makes use of the next key elements:

• SageMaker Unified Studio – Gives the event surroundings
• Circulate app – Orchestrates the workflow, together with:
  • Information base queries
  • Immediate-based classification
  • Conditional routing
  • Agent-based response technology

The workflow processes consumer queries by the next steps:

1. A consumer submits a complaint-related query.
2. The data base supplies related grievance data.
3. The immediate classifies if the question is about decision timing.
4. Based mostly on the classification utilizing the situation, the applying takes the next motion:
   1. Routes the question to an AI agent for particular decision responses.
   2. Returns basic grievance data.
5. The applying generates an acceptable response for the consumer.

Conditions

For this instance, you want the next:

• Entry to SageMaker Unified Studio. (You have to the SageMaker Unified Studio portal URL out of your administrator). You may authenticate utilizing both:
• The IAM consumer or IAM Id Heart consumer should have acceptable permissions for:
  • SageMaker Unified Studio.
  • Amazon Bedrock (together with Amazon Bedrock Flows, Amazon Bedrock Brokers, Amazon Bedrock Immediate Administration, and Amazon Bedrock Information Bases).
  • For extra data, check with Id-based coverage examples.
• Entry to Amazon Bedrock FMs (make sure that these are enabled on your account), for instance:Anthropic’s Claude 3 Haiku (for the agent).
• Configure entry to your Amazon Bedrock serverless fashions for Amazon Bedrock in SageMaker Unified Studio tasks.
• Amazon Titan Embedding (for the data base).
• Pattern grievance information ready in CSV format for creating the data base.

Put together your information

We have now created a pattern dataset to make use of for Amazon Bedrock Information Bases. This dataset has data of complaints obtained by customer support representatives and backbone data.The next is an instance from the pattern dataset:

complaint_id,product,sub_product,situation,sub_issue,complaint_summary,action_taken,next_steps,financial_institution,state,submitted_via,resolution_type,timely_response
FIN-2024-001,04/26/24,"Mortgage","Standard mortgage","Fee situation","Escrow dispute","Buyer disputes mortgage fee improve after current escrow evaluation","Reviewed escrow evaluation, defined property tax improve influence, offered detailed fee breakdown","1. Ship written rationalization of escrow evaluation 2. Schedule annual escrow evaluation 3. Present fee help choices","Monetary Establishment-1","TX","Internet","Closed with rationalization","Sure"
FIN-2024-002,04/26/24,"Cash switch","Wire switch","Processing delay","Worldwide switch","Wire switch of $10,000 delayed, buyer involved about worldwide fee deadline","Situated wire switch in system, expedited processing, waived wire payment","1. Verify receipt with receiving financial institution 2. Replace buyer on supply 3. Doc course of enchancment wants","Monetary Establishment-2","FL","Cellphone","Closed with financial reduction","No"

Create a venture

In SageMaker Unified Studio, customers can use tasks to collaborate on varied enterprise use instances. Inside tasks, you’ll be able to handle information property within the SageMaker Unified Studio catalog, carry out information evaluation, arrange workflows, develop ML fashions, construct generative AI functions, and extra.

To create a venture, full the next steps:

1. Open the SageMaker Unified Studio touchdown web page utilizing the URL out of your admin.
2. Select Create venture.
3. Enter a venture title and non-compulsory description.
4. For Undertaking profile, select Generative AI utility improvement.
5. Select Proceed.

6. Full your venture configuration, then select Create venture.

Create a immediate

Let’s create a reusable immediate to seize the directions for FMs, which we’ll use later whereas creating the circulate utility. For extra data, see Reuse and share Amazon Bedrock prompts.

1. In SageMaker Unified Studio, on the Construct menu, select Immediate underneath Machine Studying & Generative AI.

2. Present a reputation for the immediate.
3. Select the suitable FM (for this instance, we select Claude 3 Haiku).
4. For Immediate message, we enter the next:

You're a grievance evaluation classifier. You'll obtain grievance information from a data base. Analyze the {{enter}} and reply with a single letter:
T: If the enter accommodates details about grievance decision timing, response time, or processing timeline (whether or not well timed or delayed)
F: For all different forms of grievance data
Return solely 'T' or 'F' primarily based on whether or not the data base response is about decision timing. Don't add any extra textual content or rationalization - reply with simply the only letter 'T' or 'F'.

5. Select Save.

6. Select Create model.

Create a chat agent

Let’s create a chat agent to deal with particular decision responses. Full the next steps:

1. In SageMaker Unified Studio, on the Construct menu, select Chat agent underneath Machine Studying & Generative AI.
2. Present a reputation for the immediate.
3. Select the suitable FM (for this instance, we select Claude 3 Haiku).
4. For Enter a system immediate, we enter the next:

You're a Monetary Complaints Assistant AI. You'll obtain grievance data from a data base and questions on decision timing.
When responding to decision timing queries:
1. Use the offered grievance data to substantiate if it was resolved inside timeline
2. For well timed resolutions, present:
   - Affirmation of well timed completion
   - Particular actions taken (from the offered grievance information)
   - Subsequent steps that had been accomplished
2. For delayed resolutions, present:
   - Acknowledgment of delay
   - Normal compensation bundle:
     • $75 service credit score
     • Precedence Standing improve for six months
     • Service charges waived for present billing cycle
   - Actions taken (from the offered grievance information)
   - Contact data for follow-up: Precedence Line: ************** 
All the time reference the precise grievance particulars offered in your enter when discussing actions taken and backbone course of.

5. Select Save.

6. After the agent is saved, select Deploy.
7. For Alias title, enter demoAlias.
8. Select Deploy.

Create a circulate

Now that we’ve our immediate and agent prepared, let’s create a circulate that may orchestrate the grievance dealing with course of:

1. In SageMaker Unified Studio, on the Construct menu, select Circulate underneath Machine Studying & Generative AI.

2. Create a brand new circulate known as demo-flow.

Add a data base to your circulate utility

Full the next steps so as to add a data base node to the circulate:

1. Within the navigation pane, on the Nodes tab, select Information Base.
2. On the Configure tab, present the next data:
   1. For Node title, enter a reputation (for instance, complaints_kb).
   2. Select Create new Information Base.
3. Within the Create Information Base pane, enter the next data:
   1. For Identify, enter a reputation (for instance, complaints).
   2. For Description, enter an outline (for instance, consumer complaints data).
   3. For Add information sources, choose Native file and add the complaints.txt file.
   4. For Embeddings mannequin, select Titan Textual content Embeddings V2.
   5. For Vector retailer, select OpenSearch Serverless.
   6. Select Create.

4. After you create the data base, select it within the circulate.
5. Within the particulars title, present the next data:
6. For Response technology mannequin, select Claude 3 Haiku.
7. Join the output of the circulate enter node with the enter of the data base node.
8. Join the output of the data base node with the enter of the circulate output node.

9. Select Save.

Add a immediate to your circulate utility

Now let’s add the immediate you created earlier to the circulate:

1. On the Nodes tab within the Circulate app builder pane, add a immediate node.
2. On the Configure tab for the immediate node, present the next data:
3. For Node title, enter a reputation (for instance, demo_prompt).
4. For Immediate, select financeAssistantPrompt.
5. For Model, select 1.
6. Join the output of the data base node with the enter of the immediate node.
   
7. Select Save.

Add a situation to your circulate utility

The situation node determines how the circulate handles several types of queries. It evaluates whether or not a question is about decision timing or basic grievance data, enabling the circulate to route the question appropriately. When a question is about decision timing, it is going to be directed to the chat agent for specialised dealing with; in any other case, it’s going to obtain a direct response from the data base. Full the next steps so as to add a situation:

1. On the Nodes tab within the Circulate app builder pane, add a situation node.
2. On the Configure tab for the situation node, present the next data:
   1. For Node title, enter a reputation (for instance, demo_condition).
   2. Beneath Circumstances, for Situation, enter conditionInput == "T".
   3. Join the output of the immediate node with the enter of the situation node.
      
3. Select Save.

Add a chat agent to your circulate utility

Now let’s add the chat agent you created earlier to the circulate:

1. On the Nodes tab within the Circulate app builder pane, add the agent node.
2. On the Configure tab for the agent node, present the next data:
   1. For Node title, enter a reputation (for instance, demo_agent).
   2. For Chat agent, select DemoAgent.
   3. For Alias, select demoAlias.
3. Create the next node connections:
   1. Join the enter of the situation node (demo_condition) to the output of the immediate node (demo_prompt).
   2. Join the output of the situation node:
      1. Set If situation is true to the agent node (demo_agent).
      2. Set If situation is fake to the prevailing circulate output node (FlowOutputNode).
   3. Join the output of the data base node (complaints_kb) to the enter of the next:
      1. The agent node (demo_agent).
      2. The circulate output node (FlowOutputNode).
   4. Join the output of the agent node (demo_agent) to a brand new circulate output node named FlowOutputNode_2.
      
4. Select Save.

Check the circulate utility

Now that the circulate utility is prepared, let’s take a look at it. On the fitting facet of the web page, select the increase icon to open the Check pane.

Within the Enter immediate textual content field, we will ask a couple of questions associated to the dataset created earlier. The next screenshots present some examples.

Clear up

To wash up your assets, delete the circulate, agent, immediate, data base, and related OpenSearch Serverless assets.

Conclusion

On this submit, we demonstrated the right way to construct an AI-powered grievance reference system utilizing a circulate utility in SageMaker Unified Studio. Through the use of the built-in capabilities of SageMaker Unified Studio with Amazon Bedrock options like Amazon Bedrock Information Bases, Amazon Bedrock Brokers, and Amazon Bedrock Flows, you’ll be able to quickly develop and deploy refined AI functions with out intensive coding.

As you construct AI workflows utilizing SageMaker Unified Studio, keep in mind to stick to the AWS Shared Duty Mannequin for safety. Implement SageMaker Unified Studio safety finest practices, together with correct IAM configurations and information encryption. You can too check with Safe a generative AI assistant with OWASP Prime 10 mitigation for particulars on the right way to assess the safety posture of a generative AI assistant utilizing OWASP TOP 10 mitigations for frequent threats. Following these pointers helps set up sturdy AI functions that keep information integrity and system safety.

To be taught extra, check with Amazon Bedrock in SageMaker Unified Studio and be part of discussions and share your experiences in AWS Generative AI Community.

We look ahead to seeing the modern options you’ll create with these highly effective new options.

Concerning the authors

Sumeet Tripathi is an Enterprise Assist Lead (TAM) at AWS in North Carolina. He has over 17 years of expertise in know-how throughout varied roles. He’s keen about serving to clients to cut back operational challenges and friction. His focus space is AI/ML and Power & Utilities Phase. Exterior work, He enjoys touring with household, watching cricket and films.

Vishal Naik is a Sr. Options Architect at Amazon Internet Providers (AWS). He’s a builder who enjoys serving to clients accomplish their enterprise wants and clear up complicated challenges with AWS options and finest practices. His core space of focus consists of Generative AI and Machine Studying. In his spare time, Vishal loves making brief movies on time journey and alternate universe themes.