Amazon Bedrock Managed Knowledge Base can now answer questions that require calculations across complete CSV or Excel files. Its tabular search feature retrieves a relevant table, stages the original file in an isolated environment with no network access, and runs model-written code over every needed row or sheet.
Conventional retrieval-augmented generation splits documents into passages. That works for prose but can omit rows from long tables or columns from very wide ones, producing incomplete totals and comparisons. Tabular search instead uses retrieval to find the right file, then lets code perform lookups, arithmetic and threshold checks on the whole table. The final answer can combine those results with facts and citations from related policy documents.
The capability works through the existing AgenticRetrieveStream request and supports CSV, XLSX and XLS files brought in through connectors including Amazon S3, Microsoft SharePoint and Atlassian Confluence. A trace identifies the table the generated code used, while the answer receives only what that code prints.
Running code adds at least one agent turn, so table-based responses take longer than text-only answers. AWS also cautions that governed metrics, recurring dashboards, warehouse-scale joins and formal reporting controls still belong in business-intelligence systems rather than this schema-free search tool.