Create a journey planning agentic workflow with Amazon Nova

Traveling is pleasing, however journey planning may be advanced to navigate and a problem. Travelers should guide lodging, plan actions, and prepare native transportation. All these selections can really feel overwhelming. Although journey professionals have lengthy helped handle these complexities, latest breakthroughs in generative AI have made one thing totally new potential—clever assistants that may perceive pure dialog, entry real-time information, and instantly interface with reserving techniques and journey instruments. Agentic workflows, which use massive language fashions (LLMs) with entry to exterior instruments, are significantly promising for simplifying dynamic, multi-step processes like journey planning.

In this submit, we discover learn how to construct a journey planning resolution utilizing AI brokers. The agent makes use of Amazon Nova, which gives an optimum stability of efficiency and value in comparison with different business LLMs. By combining correct however cost-efficient Amazon Nova fashions with LangGraph orchestration capabilities, we create a sensible journey assistant that may deal with advanced planning duties whereas protecting operational prices manageable for manufacturing deployments.

Solution overview

Our resolution is constructed on a serverless AWS Lambda structure utilizing Docker containers and implements a complete three-layer strategy: frontend interplay, core processing, and integration companies. In the core processing layer, we use LangGraph, a stateful orchestration framework, to create a complicated but versatile agent-based system that manages the advanced interactions required for journey planning.

The core of our system is a graph structure the place parts (nodes) deal with distinct elements of journey planning, with the router node orchestrating the stream of knowledge between them. We use Amazon Nova, a brand new technology of state-of-the-art basis fashions (FMs) out there solely on Amazon Bedrock that delivers frontier intelligence with industry-leading price-performance. The router node makes use of an LLM to research every consumer question and, with entry to the outline of our 14 motion nodes, decides which of them should be executed. The motion nodes, every with their very own LLM chain, powered by both Amazon Nova Pro or Amazon Nova Lite fashions, handle numerous capabilities, together with internet analysis, customized suggestions, climate lookups, product searches, and procuring cart administration.

We use Amazon Nova Lite for the router and less complicated motion nodes. It can deal with question evaluation and fundamental content material technology with its lightning-fast processing whereas sustaining sturdy accuracy at a low price. Five advanced nodes use Amazon Nova Pro for duties requiring superior instruction following and multi-step operations, akin to detailed journey planning and suggestions. Both fashions assist a 300,000-token context window and may course of textual content, picture, and video inputs. The fashions assist textual content processing throughout greater than 200 languages, serving to our journey assistant serve a world viewers.The integration layer unifies a number of information sources and companies by an interface:

These integrations function examples, and the structure is designed to be extensible, so organizations can shortly incorporate their very own APIs and information sources based mostly on particular necessities.

The agent retains monitor of the dialog state utilizing AgentState (TypedDict), a particular Python dictionary that helps forestall information errors by imposing particular information sorts. It shops the knowledge we have to learn about every consumer’s session: their dialog historical past, profile info, processing standing, and last outputs. This makes positive the totally different motion nodes can entry and replace info reliably.

The following diagram illustrates the answer structure.

The journey assistant processes consumer interactions from finish to finish:

1. Users work together with a React.js internet utility by a chat interface.
2. Their requests are authenticated utilizing Amazon Cognito and routed by Amazon API Gateway.
3. Authenticated requests are despatched to our backend Lambda capabilities, which host the core agent workflow.
4. API credentials are securely saved utilizing AWS Secrets Manager, following finest practices to verify these delicate keys are by no means uncovered in code or configuration information, with applicable entry controls and rotation insurance policies applied.
5. The Travel Assistant Agent itself consists of a number of interconnected parts. At the middle, the agent router analyzes incoming queries and orchestrates the workflow.
6. The agent maintains state by three DynamoDB tables that retailer dialog historical past, procuring wishlists, and consumer profiles, ensuring context is preserved throughout interactions.
7. For travel-specific information, the system makes use of a mix of Amazon Bedrock Knowledge Bases, Amazon OpenSearch Serverless, and a doc retailer in Amazon Simple Storage Service (Amazon S3). These parts work collectively to supply correct, related journey info when wanted.
8. The agent’s motion nodes deal with specialised duties by combining LLM chains with exterior APIs. When customers want product suggestions, the system connects to the Amazon Product Advertising API. For common journey info, it makes use of the Google Custom Search API, and for weather-related queries, it consults the OpenWeather API. API credentials are securely managed by Secrets Manager.
9. The system formulates complete responses based mostly on collected info, and the ultimate responses are returned to the consumer by the chat interface.

This structure helps each easy queries that may be dealt with by a single node and sophisticated multi-step interactions that require coordination throughout a number of parts. The system can scale horizontally, and new capabilities may be added by introducing further motion nodes and API integrations.

You can deploy this resolution utilizing the AWS Cloud Development Kit (AWS CDK), which generates an AWS CloudFormation template that handles the mandatory sources, together with Lambda capabilities, DynamoDB tables, and API configurations. The deployment creates the required AWS sources and outputs the API endpoint URL in your frontend utility.

Prerequisites

For this walkthrough, you need to have the next conditions:

Clone the repository

Start by cloning the GitHub repository containing the answer information:

git clone 

Obtain API keys

The resolution requires API keys from three companies to allow its core functionalities:

• OpenWeather API – Create a Free Access account at OpenWeather to acquire your API key. The free tier (60 calls per minute) is ample for testing and growth.
• Google Custom Search API – Set up the search performance by Google Cloud Console. Create or choose a challenge and allow the Custom Search API. Then, generate an API key from the credentials part. Create a search engine at Programmable Search and observe your Search Engine ID. The free tier consists of 100 queries per day.
• (Optional) Amazon Product Advertising API (PAAPI) – If you need to allow product suggestions, entry the PAAPI Documentation Portal to generate your API keys. You will obtain each a public key and a secret key. You should have an Amazon Associates account to entry these credentials. If you’re new to the Amazon Associates Program, full the applying course of first. Skip this step in case you don’t need to use PAAPI options.

Add API keys to Secrets Manager

Before deploying the answer, you need to securely retailer your API keys in Secrets Manager. The following desk lists the secrets and techniques to create and their JSON construction. For directions to create a secret, seek advice from Create an AWS Secrets Manager secret.

Configure setting variables

Create a .env file within the challenge root together with your configuration:

STACK_NAME=TravelAssistantAgent

# Optional: Create Bedrock Knowledge Base with paperwork
KB_DOCS_PATH = Path/to/your/paperwork/folder
# Optional: Enable/disable Product Search options with PAAPI
USE_PAAPI=false

Deploy the stack

If that is your first time utilizing the AWS CDK in your AWS account and AWS Region, bootstrap your setting:

Deploy the answer utilizing the supplied script, which creates the required AWS sources, together with Lambda capabilities, DynamoDB tables, and API configurations:

Access your utility

When the deployment is full, open the AWS CloudFormation console and open your stack. On the Outputs tab, observe the next values:

• WebAppDomain – Your utility’s URL
• UserPoolId – Required for consumer administration
• UserPoolClientId – Used for authentication

Create an Amazon Cognito consumer

Complete the next steps to create an Amazon Cognito consumer:

1. On the Amazon Cognito console, select User swimming pools within the navigation pane.
2. Choose your consumer pool.
3. Choose Users within the navigation pane, then select Create consumer.

4. For Email deal with, enter an e mail deal with, and choose Mark e mail deal with as verified.
5. For Password, enter a short lived password.
6. Choose Create consumer.

You can use these credentials to entry your utility on the WebAppDomain URL.

Test the answer

To take a look at the agent’s capabilities, we created a enterprise traveler persona and simulated a typical journey planning dialog stream. We centered on routing, perform calling accuracy, response high quality, and latency metrics. The agent’s routing system directs the consumer inquiries to the suitable specialised node (for instance, trying to find lodging, checking climate circumstances, or suggesting journey merchandise). Throughout the dialog, the agent maintains the context of beforehand mentioned particulars, so it may construct upon earlier responses whereas offering related new info. For instance, after discussing journey vacation spot, the agent can naturally incorporate this into subsequent climate and packing record suggestions.

The following screenshots reveal the end-user expertise, whereas the underlying API interactions are dealt with seamlessly on the backend. The full implementation particulars, together with Lambda perform code and API integration patterns, can be found in our GitHub repository.

The resolution demonstrates personalization capabilities utilizing pattern consumer profiles saved in DynamoDB, containing upcoming journeys and journey preferences. In manufacturing deployments, these profiles may be built-in with current buyer databases and reservation techniques to supply a personalised help.

The product suggestions proven are stay hyperlinks to precise objects out there on Amazon.com, so the consumer can discover or buy these merchandise instantly. The consumer can select a hyperlink to take a look at the product, or select Add to Amazon Cart to see the objects of their procuring cart.

Clean up

After you’re accomplished experimenting with the journey assistant, you’ll be able to find the CloudFormation stack on the AWS CloudFormation console and delete it. This will delete the sources you created.

Conclusion

Our journey planning assistant agent demonstrates a sensible utility constructed by Amazon Nova and LangGraph for fixing real-world enterprise challenges. The system streamlines advanced journey planning whereas naturally integrating product suggestions by specialised processing nodes and real-time information integration. Amazon Nova Lite fashions confirmed cheap efficiency at activity orchestration, and Amazon Nova Pro carried out nicely for extra advanced perform calling operations. Looking forward, this framework may very well be applied with extra dynamic orchestration techniques akin to ReAct. To construct your personal implementation, discover our code samples within the GitHub repository.

For these seeking to deepen their understanding of LLM-powered brokers, AWS supplies intensive sources on constructing clever techniques. The Amazon Bedrock Agents documentation gives insights into automating multistep duties with FMs, and the AWS Bedrock Agent Samples GitHub repo supplies steering for implementing a number of agent functions utilizing Amazon Bedrock.

About the authors

Isaac Privitera is a Principal Data Scientist with the AWS Generative AI Innovation Center, the place he develops bespoke generative AI-based options to deal with clients’ enterprise issues. His major focus lies in constructing accountable AI techniques, utilizing strategies akin to RAG, multi-agent techniques, and mannequin fine-tuning. When not immersed on this planet of AI, Isaac may be discovered on the golf course, having fun with a soccer recreation, or climbing trails along with his loyal canine companion, Barry.

Ryan Razkenari is a Deep Learning Architect on the AWS Generative AI Innovation Center, the place he makes use of his experience to create cutting-edge AI options. With a powerful background in AI and analytics, he’s obsessed with constructing modern applied sciences that deal with real-world challenges for AWS clients.

Sungmin Hong is a Senior Applied Scientist on the AWS Generative AI Innovation Center, the place he helps expedite a wide range of use circumstances for AWS clients. Before becoming a member of Amazon, Sungmin was a postdoctoral analysis fellow at Harvard Medical School. He holds a PhD in Computer Science from New York University. Outside of labor, Sungmin enjoys climbing, studying, and cooking.