Amazon Bedrock Can Now Count Across Whole Excel Tables, Not Just Search Snippets
Standard RAG chops a table into chunks, so the model sees only a few rows and the totals come out incomplete. Amazon Bedrock Managed Knowledge Base now adds tabular search: code running in a sandbox with no network access reads the entire CSV or Excel file. In AWS's example, it found a department 124,559.90 dollars over budget and cited the company policy. For questions that combine a rule from a document with a calculation over a table, one query is now enough.
AI-processed from AWS Machine Learning Blog; edited by Hamidun News
Picture a finance team at the end of the quarter. It needs to determine which departments must submit explanations for budget variances. The answer is not in any single file: the variance threshold is written in the policy, the revised budgets sit in a quarter-close memo, and the figures are buried in an Excel workbook of profit and loss data. Ordinary document search handles this task poorly. AWS has added tabular search to Amazon Bedrock Managed Knowledge Base, targeting exactly this gap between text and tables.
Why tables break ordinary RAG
RAG returns a fixed number of chunks that best match the question. For prose, this works. For tables, it works worse: when a file is split into pieces, some of the information is lost.
If a table is long, with thousands of rows, search returns only a few chunks. A total, a comparison or a threshold check is then built on a handful of rows and comes out incomplete. If a table is wide, with hundreds of columns, a single row may not fit into a chunk or into the search model's window. Values from distant columns are harder to retrieve, and long chunks start matching unrelated questions and crowd out useful ones.
There is also a matching problem. People ask about "controllable expenses," while the table contains cost center codes and account names. A draft and a final version of a workbook look almost identical to search. And the rule that gives the numbers their meaning often sits in a separate PDF.
How it works
Tabular search splits the job into two parts. Search only has to find one chunk from the right tabular file; after that, code reads the table in full. Nothing needs to be enabled: the feature is built into agentic retrieval, that is, into the AgenticRetrieveStream request.
A request goes through several stages:
- Search returns chunks, among which is a piece from a CSV, XLSX or XLS file;
- Based on the question and the table's structure, the agent decides whether code needs to run;
- The source file is copied into an isolated sandbox with no network access;
- Code written by the model calculates over the entire table or sheet;
- The agent receives only the printed result, such as totals or matching rows;
- The answer is assembled from the calculation result and document chunks, with links to sources.
In AWS's example, the company AnyCompany Retail asks which cost center had the largest overspend on controllable expenses in the second fiscal quarter, in the sense of the financial policy. From the PDF, the agent takes the definition: it includes contractors, travel, software and marketing programs, and excludes salaries and rent. From the tables, the code takes all rows from the Actuals and Budget sheets, filters them by quarter and four accounts, and compares cost centers.
The result: cost center CC-3010 (Platform) exceeded its budget by 124,559.90 dollars, spending 394,859.90 against a plan of 270,300.
00. The business review explains the cause: contractors for a data platform migration. If all operating expenses are counted rather than only controllable ones, the figure is 133,434.
05.
Tracing is also useful. The event stream shows which file the code read: a Planning event lists the documents, followed by CodeExecution events with the statuses IN_PROGRESS and SUCCEEDED. This makes it possible to verify that the agent used the final workbook rather than a draft. The links in the answer point to the chunks used, so the table itself may not appear in the citations, but it will be visible in the trace.
Limitations and what to do
The feature does not replace BI. Running code adds at least one agent turn, so answers involving tabular data arrive more slowly than text-only ones. Verified metrics, regular dashboards, warehouse-scale joins and reporting controls should still be handled in BI tools. In addition, search must still find the right file first, so it is worth naming the workbook explicitly in the question.
There are also practical requirements. The calling account needs the bedrock:AgenticRetrieveStream, bedrock:Retrieve and bedrock:GetDocumentContent permissions, because the agent reads the source file with its rights. After the first sync, check that tabular chunks have the file type CSV or MS_EXCEL; otherwise, tabular search will not run. A regular Retrieve call still returns only chunks and does not execute code, as does the mode with generateResponse=false.
For developers, the takeaway is simple. If your assistant answers questions where the rule lives in a document and the numbers live in exports and workbooks, you do not need to build a schema in advance or load files into a warehouse. Put the tables next to the PDFs in S3 or SharePoint, phrase questions with the file name, and check the trace. For one-off analytical questions about working files, this may be enough. Anything that must be calculated regularly and with guarantees stays in BI.
- Nov 21, 2026Amazon Bedrock Adds Explicit Prompt Caching for Open-Weight Models
- Nov 21, 2026Huang Steps Down at Nvidia as AMD's Share of AI Chips Grows from 2% to 7–10%
- Nov 21, 2026AssemblyAI cut GPU inference costs by 70% by moving to AWS Spot Instances
- Nov 21, 2026AWS shows how to cap Bedrock spending per employee without writing code
Key Questions Answered
Why tables break ordinary RAG?
RAG returns a fixed number of chunks that best match the question. For prose, this works. For tables, it works worse: when a file is split into pieces, some of the information is lost. If a table is long, with thousands of rows, search returns only a few chunks. A total, a comparison or a threshold check is then built on a handful of rows and comes out incomplete. If a table is wide, with hundreds of columns, a single row may not fit into a chunk or into the search model's window.
How it works?
Tabular search splits the job into two parts. Search only has to find one chunk from the right tabular file; after that, code reads the table in full. Nothing needs to be enabled: the feature is built into agentic retrieval, that is, into the AgenticRetrieveStream request.
What are the key details on Limitations and what to do?
The feature does not replace BI. Running code adds at least one agent turn, so answers involving tabular data arrive more slowly than text-only ones. Verified metrics, regular dashboards, warehouse-scale joins and reporting controls should still be handled in BI tools. In addition, search must still find the right file first, so it is worth naming the workbook explicitly in the question. There are also practical requirements.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.