Amazon Bedrock Managed Knowledge Base has been updated to support tabular search, a capability that automatically loads full CSV, XLSX, or XLS files into an isolated sandbox, runs model‑generated code over the entire table, and merges the results with retrieved document passages. The change is transparent to existing AgenticRetrieveStream calls, but it expands the type of questions that can be answered directly from a knowledge base.
Why the addition matters to practitioners
Engineers building generative‑AI assistants often hit a wall when a query requires aggregating rows or performing look‑ups across large spreadsheets. Traditional retrieval returns only a handful of passages, which can omit critical rows and prevent calculations such as totals, percentages, or threshold checks. With tabular search, the retrieval step only needs to locate the correct file; the subsequent sandbox execution handles the full dataset, delivering complete analytical answers and proper citations.
Architectural and operational implications
From an architecture perspective, the workflow now includes an extra stage after the initial retrieval:
- Retrieve: Passages are returned, including a reference to a CSV/XLSX file.
- Plan: The agent decides whether the question requires table processing.
- Stage: The original file is copied into a sandbox with no external network access.
- Compute: Model‑generated code runs over the whole table, emitting only the values needed for the answer.
- Return: The agent receives the computed output and combines it with document passages.
- Answer: The final response includes citations to both the policy documents and the workbook that was read.
Implementation does not require new API parameters; the same AgenticRetrieveStream request triggers tabular search when the retrieved passage points to a supported file type and the question implies a row‑wise operation. Practitioners should ensure that relevant tables are stored in accessible locations such as Amazon S3 or SharePoint, and that file names are referenced in queries to improve retrieval accuracy.
Security considerations
The sandbox used for tabular execution is isolated and has no outbound network connectivity, limiting exposure to external services. This design reduces the attack surface compared to running custom extraction code in a shared environment. However, because the sandbox reads the entire source file, engineers must verify that the file does not contain sensitive data that should not be exposed to the model. Auditing the trace output— which names the workbook that was read—provides visibility into which assets are being accessed during a request.
Related CloudNinjas coverage: AWS.
What This Means For Practitioners
Adopt tabular search when queries span both structured tables and unstructured documents, and replace ad‑hoc ETL scripts with the built‑in sandbox execution. Validate that your storage locations and naming conventions support reliable file discovery, and monitor sandbox traces to confirm that only intended files are processed. For security‑focused teams, treat the sandbox as a controlled compute boundary and incorporate its usage into data‑handling policies.


