Skip to content
AI.info

The Pulse

Amazon Bedrock adds search across spreadsheet tables

Amazon Bedrock Managed Knowledge Base can now calculate and look up values across spreadsheet tables during agentic retrieval. The feature combines results from CSV and Excel files with information from related documents.

Amazon Bedrock adds search across spreadsheet tables

Amazon Bedrock Managed Knowledge Base can now calculate and look up values across spreadsheet tables when answering questions through agentic retrieval. The feature works with CSV, XLSX and XLS files and can combine table results with information from related documents, AWS said in a post on November 16.

From passages to whole tables

When a question calls for table data, the retrieval process first finds passages from relevant files. An agent then determines whether to run code over the retrieved tables; if it does, the original files are copied into an isolated sandbox without network access.

Model-written code reads the whole table or sheet, and the agent receives what the code prints, such as totals or matching rows. It can then combine those results with retrieved document passages in an answer with citations. A trace identifies the table files the code read.

Tabular search is part of agentic retrieval and does not require a separate switch. An AgenticRetrieveStream request can search one knowledge base or several. The separate Retrieve API continues to return passages without running code.

Why the file matters

A conventional retrieval system returns a limited set of matching passages rather than the entirety of a long table. AWS says that can leave a total, comparison or threshold check incomplete if the answer relies on only the retrieved rows.

Wide tables pose a different problem: a row can contain more text than a passage or search model can take in at once, making later columns harder to match. Tabular search divides the work between finding a passage from the right file and then reading the whole table with code.

Retrieval still has to find the relevant file before the calculation can happen. AWS says naming that file in the question helps.

A finance example

AWS illustrates the feature with fictitious finance files, including a final profit-and-loss workbook, a draft workbook, a variance policy and a quarterly close memo. The example asks which departments must write variance commentary, a question that depends on a policy threshold, budget revisions and calculations across departments.

According to the example, agentic retrieval with tabular search identifies Sales, Engineering and Operations and gives each department’s variance against its revised budget. The answer cites the policy and memo, while the trace names the workbook read by the code. In the comparison shown by AWS, retrieval alone returns passages but no complete answer.

Access and limits

The feature works with files brought into a knowledge base through connectors including Amazon S3, Microsoft SharePoint and Atlassian Confluence. It is available in supported AWS Regions, and the caller needs agentic retrieval permissions, including bedrock:GetDocumentContent; the agent reads the original table file using the caller’s credentials.

AWS says answers using table data take longer than text-only answers because running code adds at least one agent turn. The company positions tabular search for calculations and lookups over retrieved files, while directing governed metrics, recurring dashboards, warehouse-scale joins and reporting controls to business intelligence tools.

Sources

Companies in this story

Related stories

Explore

More articles