Shoppers ask with typos, half-sentences
and "how long does that take?"

Two questions that mean the same thing rarely use the same words, which is why the understanding layer works on meaning rather than on the text itself.

• Amazon Bedrock supplies the foundation models that interpret queries and generate responses, and models can be upgraded as new ones ship without rebuilding the pipeline.
• Amazon Titan embeddings match by meaning, so paraphrases land on the right intent instead of depending on keyword overlap.
• Multi-turn conversation handling keeps context across exchanges, so "how long does that take?" refers back to the return being discussed.

Each merchant's knowledge sits in different places, updated on different schedules. We brought it into one retrieval layer, and the model generates only from what it retrieves:

• Amazon Kendra pulls FAQs, policy documents and product catalogs into the searchable knowledge layer.
• Amazon OpenSearch handles vector retrieval of the relevant source material, which the retrieval-augmented generation (RAG) pipeline passes into generation.
• AWS Lambda orchestrates retrieval, generation and integration logic.
• Index and embedding work targeted sub-second retrieval during peak query volume. The target is for retrieval, not including response generation.

The ingestion pipeline lets authorized merchant updates enter the knowledge layer without an application redeployment, and they reach the AI agent within minutes. That matters most for the policies and product details that change during a promotion.

Why: the model handles language well but doesn't know facts it hasn't been given, so retrieval feeds it those facts and the model turns them into an answer.

An AI agent that can't learn will still get a question wrong on day 90 if it got it wrong on day 1. We capture feedback on each response so it doesn't stay buried in the conversation history.

• Shopper ratings, support-staff corrections and escalation outcomes are stored with the interaction's context.
• An Amazon SageMaker pipeline aggregates that feedback, evaluates model performance and surfaces candidates for refinement.
• Reporting separates accuracy and resolution by merchant, so the team can investigate performance against a particular catalog and knowledge base, and show each merchant how the AI agent does on its own customers.

Black Friday and merchant promotions multiply support volume within hours, and an AI agent that slows down at peak causes real problems for shoppers trying to get help. The runtime uses containers on Amazon ECS/Fargate with autoscaling so capacity follows demand. Amazon RDS holds transactional state with merchant-level isolation, which means one merchant's shoppers and conversations can never reach another's. Amazon S3 stores conversation history, feedback and audit data, with lifecycle policies for retention.

Encryption and access controls protect the stored data and the service interfaces, CloudWatch tracks latency and availability, IAM sets the access boundaries and AWS WAF protects the public traffic path. Because security is built in this way, a merchant's security review gets its answers straight from the architecture.