Enterprise answers often span cells and prose

Amazon Bedrock Managed Knowledge Base now supports tabular search over CSV, XLSX and XLS files, allowing agentic retrieval to calculate across rows and combine the result with evidence from ordinary documents. AWS is targeting questions that cannot be answered by finding one similar paragraph. A quarterly variance decision, for example, may require a policy threshold from a memo, revised budget figures from another document and arithmetic across every department in a spreadsheet. The feature brings those steps into one managed retrieval workflow.

Traditional vector search performs well when the answer resembles a passage that can be embedded and ranked. Tables are different. Meaning depends on headers, units, row relationships and calculations over a complete set. Splitting a workbook into text chunks can separate a number from its label or omit rows required for an aggregate. Tabular search recognises that analytical retrieval needs structured operations rather than semantic similarity alone, then lets the agent combine calculated values with unstructured context.

The supported formats cover common business exports and spreadsheets. A knowledge base can index one or more tables or sheets, identify relevant rows and perform multi-step calculations before generating an answer. That could support finance close, inventory analysis, operational reporting and policy checks. It is not a replacement for a governed warehouse or semantic layer. Large, frequently changing datasets and complex joins may still belong in systems designed for transactional consistency and reproducible queries.

Tabular search adds calculation to retrieval

Schema quality will determine whether the feature produces sensible results. Merged cells, repeated headers, hidden rows, inconsistent date formats and numbers stored as text can confuse an automated parser. Units such as thousands, percentages or currencies need explicit labels. Teams should create a controlled ingestion profile that validates workbooks, identifies the authoritative sheet and rejects ambiguous layouts. Quietly accepting a malformed file is more dangerous than failing, because the resulting calculation can look precise while using the wrong rows.

Calculation also introduces a verification requirement. Answers should expose the source files, filters, columns and formula or operation used, not only a final number. A reviewer needs to distinguish a sum from an average, understand whether blanks were ignored and see which period or business unit was selected. Golden test questions with independently calculated results can detect regressions when the retrieval service, source workbook or instruction changes. High-impact financial outputs should remain subject to established review and approval.

Data preparation determines reliability

Combining tables with documents is the feature’s most useful aspect. A policy may define which threshold matters, while the spreadsheet contains the values to test. The agent can retrieve both and explain why a row meets a condition. This reduces the temptation to put prose into spreadsheet cells solely for search. It also creates more ways to be wrong: the system may select an obsolete policy, an unauthorised workbook or mismatched reporting periods. Metadata and lifecycle state should be first-class filters.

Access control must reach individual data sources and, where required, rows or tenants. A conversational interface can make broad discovery easier than the underlying file systems intended. Identity from the calling application should propagate into retrieval, and cached or generated answers must not outlive the user’s permission. Sensitive sheets should be separated rather than relying on prompts to hide columns. Audit logs should record who asked, which sources were accessed and what calculation was returned.

Access and evidence must survive the abstraction

Operational teams also need freshness guarantees. A workbook uploaded last week may no longer represent the close being discussed today. Ingestion pipelines should expose source timestamps, checksum changes and indexing status. Answers should say when a relevant table is absent or stale rather than extrapolating from older data. A visible citation to an outdated file does not make the result current. Owners should define the maximum acceptable age for each dataset and alert when it is exceeded.

Tabular search expands Bedrock Managed Knowledge Base from document retrieval towards mixed analytical reasoning. That can remove brittle custom code for bounded business questions and give users a more natural interface to existing files. Australian organisations should pilot it on low-risk, well-structured datasets, compare every answer with a deterministic calculation and test permission boundaries. The capability is most trustworthy when it preserves the table’s structure, shows its working and refuses to invent a result when the necessary rows or policy documents are unavailable.

Evaluation should include questions with no valid answer, conflicting versions of a table and deliberately malformed cells. A useful system must abstain or explain uncertainty instead of calculating confidently over the wrong range. Teams should retain the selected rows, formula or execution trace and cited documents with each answer. Those artefacts let analysts reproduce the result and distinguish a retrieval error from a calculation or source-data error. Production acceptance should set numeric tolerances, citation requirements and escalation rules before business users depend on the answers.