Plain retrieval returns a fixed number of passages, so answers built from long or wide tables can miss the rows needed for totals, comparisons or threshold checks. With tabular search, the agent retrieves passages, then decides whether to run model-written code over one or more retrieved table files. The code reads the original file, and the agent receives only what the code prints.
The feature needs no separate setup and works through the existing AgenticRetrieveStream request, while the Retrieve API still returns passages without running code. It works with files from connectors such as Amazon S3, Microsoft SharePoint and Atlassian Confluence. Answers can combine table calculations with rules from policies or memos stored in the same knowledge base.
Each table query adds at least one agent turn, so answers take longer than text-only ones. Governed metrics, recurring dashboards and warehouse-scale joins remain better suited to business intelligence tools. As of this writing, code runs over CSV files up to 500 MB but over Excel sheets only up to 200 MB.
Why it matters if you build agents
Agents can now answer questions that need a total, a comparison or a threshold check across every row of a table, combined with rules from policy documents. Developers should confirm that table files are CSV or Excel and fall within the file size filter, since files above maxFileSizeInMegaBytes are not ingested and the code cannot read them. The calling identity also needs bedrock:GetDocumentContent, because the agent reads the original file with the caller's credentials.
Get stories like this every weekday
One short email with the day's agent tooling news.