Couchbase built a multimodel AI architecture for Capella iQ with Amazon Bedrock

In the AWS ML blog, it is discussed how Couchbase built a multi-model AI architecture for Capella iQ, an AI developer assistant that generates database queries, recommends indexes, and supports multi-turn dialogues. The architecture is deployed across two AWS regions (us-east-1 and us-west-2) using Amazon EKS, with model invocations handled through Amazon Bedrock. Developer requests are processed by microservices cp-api, cp-internal-api, and cp-ns, which determine routing and model configurations. All traffic goes through a VPC interface, never leaving the public internet. Cross-Region Inference automatically redirects requests between US regions during failures or load spikes. For model selection, Couchbase conducted benchmarking using a BIRD-like methodology: Claude Sonnet 4.5 achieved approximately 76% accuracy and was chosen as the primary model. Amazon Bedrock made the architecture provider-agnostic: changing models requires only a configuration update, no code changes. In the future, Couchbase plans to use Bedrock Custom Model Import to optimize costs for high-frequency tasks.