Breaking02 / IT
AWS shows a semantic layer for AI agents inside EDB Postgres
The same agent code returns different results depending on the database role that runs it.

AWS published a design on 4 Nov 2026 for building a semantic layer with EDB Postgres AI. Developers who use it with Amazon Bedrock AgentCore keep data meaning and permissions in the database.
A semantic layer tells agents what data means, according to the AWS post. The AIDB extension takes business meaning from COMMENT statements and join paths from foreign keys. If an agent reads the order-total column as a unit price, top-customer spend comes out at $502,529 rather than $27,991.66.
The documentation for AIDB's MCP endpoint states that "Postgres grants decide what each agent can run". MCP, the Model Context Protocol, is a standard that connects agents to tools. In an AWS example, a role without permission on customers.ssn receives a denial error even though the code is the same.
Compatibility. The AIDB AMI, a server image, includes a default knowledge base called semkb, according to AWS. Once a new knowledge base is added, every search call has to specify kb_name.
Where it runs. Embeddings run on a local llama.cpp model in the AIDB AMI, AWS said. Claude on Amazon Bedrock handles query planning and answer writing. AgentCore Gateway connects to the MCP server without AWS Lambda.
Earlier AWS posts built the same customer-analysis case with Stardog, Denodo and others. This design places the meaning layer in the operational database. AWS said that leaves no second system to deploy and no catalogue to synchronise.
The post does not give the size of the gain in text-to-SQL accuracy from comments, which AWS describes as the largest.