AWS released a design pattern called Adjudicated Query that combines generative AI with deterministic rules engines. The architecture allows business users to query large document sets, such as apartment leases, via Amazon Q. While AI handles exploratory search, a non-AI engine makes the final compliance decisions to ensure verifiable results.
Hybrid compliance architecture
The Adjudicated Query pattern addresses limitations in standard retrieval-augmented generation and text-to-SQL approaches, which can fail to guarantee completeness or provide verifiable evidence. In this new architecture, a compliance officer uses the Amazon Q chat interface to initiate queries. A backend system then processes these requests using a combination of large language models and a deterministic rules engine.
The workflow utilizes an AWS Lambda function to host a Model Context Protocol server. While Amazon Bedrock is employed for exploratory clause searching, the actual pass or fail compliance decisions are handled by a non-AI rules engine. This separation ensures that the resulting compliance data is based on fixed logic rather than probabilistic AI outputs, which helps avoid issues like hallucinated predicates that could silently exclude relevant documents.
Availability and technical implementation
AWS has deployed a working sample and reference architecture to demonstrate the pattern using lease compliance as the primary example. The system integrates several services, including Amazon Cognito for authentication and Amazon Aurora Serverless v2 for data storage. Compliance results are surfaced through both the chat agent and an Amazon QuickSight dashboard to provide a single source of truth.
The developer notes that while the sample focuses on apartment leases, the pattern is designed for other high-stakes domains such as sanctions screening, export control, and insurance claims. The implementation requires specific AWS configurations, including VPC connections for dashboard reporting and API Gateway for handling model requests. The source provides the design pattern and code rather than a standalone automated service.
Original source
This report summarises the source below. Analysis is labelled separately; product and research claims remain attributed to their source.
Read the original at AWS ML