# yaml-language-server: $schema=../references/diagram.schema.json
title: Retrieval augmented generation on AWS
subtitle: Documents are chunked and embedded into OpenSearch Serverless; an AgentCore agent retrieves context and answers with a Bedrock model
direction: down
# merge: false gives every edge into the models, the vector index and CloudWatch its own port;
# merged trunks would hide which flow is which.
layout: { merge: false }

legend:
  solid: Requests and data
  dashed: Authentication
  dotted: Telemetry

groups:
  cloud: { type: aws:cloud }
  region: { type: aws:region, label: us-east-1, in: cloud }
  # Two lanes so the flows do not interleave; both models and the vector index sit outside them, where the flows meet.
  query: { type: zone, label: Query, in: region }
  ingestion: { type: zone, label: Ingestion, in: region }

nodes:
  users: { icon: aws:users, label: Users }
  auth: { icon: aws:cognito, label: Amazon Cognito, note: User pool, in: query }
  api: { icon: aws:api-gateway, label: Amazon API Gateway, in: query }
  agent: { icon: aws:bedrock-agentcore, label: RAG agent, note: Amazon Bedrock AgentCore Runtime, in: query }
  llm: { icon: aws:bedrock, label: Amazon Bedrock, note: Text generation model, in: region }
  docs: { icon: aws:simple-storage-service-bucket, label: Documents, note: Amazon S3 bucket, in: ingestion }
  ingest: { icon: aws:lambda-lambda-function, label: Ingestion function, note: Chunk and embed, in: ingestion }
  embed: { icon: aws:bedrock, label: Amazon Bedrock, note: Titan Text Embeddings V2, in: region }
  vectors: { icon: aws:opensearch-service-index, label: Vector index, note: Amazon OpenSearch Serverless, in: region }
  logs: { icon: aws:cloudwatch, label: Amazon CloudWatch, note: Logs and traces, in: region }

edges:
  - users --> auth: sign in
  - users -> api: question
  - { from: auth, to: api, arrow: start, line: dashed, label: verify token }
  - api -> agent
  - agent -> llm: prompt and context
  - agent -> embed: embed question
  - agent -> vectors: retrieve
  - docs -> ingest: ObjectCreated
  - ingest -> embed: chunks
  - ingest -> vectors: index
  - agent, ingest ..> logs