Beyond tables: documents and unstructured data as agent context

Most of this site is about structured data: tables, models, dashboards and the metadata around them. But the questions agents get often need context that lives elsewhere. Why was a metric redefined? Which playbook applies to this customer segment? What did the finance team decide in last quarter's review? That knowledge sits in documents, wikis and conversations.

Collibra lists Deasy Labs for unstructured data such as emails, PDFs and SharePoint on its homepage.

Source: Collibra homepage

Alation's Governed Collections, announced on 17 September 2026 and in early access at the time, bring documents from Confluence, SharePoint and S3 under governance as context for AI.

Source: Alation press release: AIOS expansion with six products

Jedify says its context graph draws on documents, playbooks, Slack and meeting recordings alongside warehouses, CRMs, financial systems and BI tools.

Source: GlobeNewswire: Jedify raises $24 million Series A

DataHub's context platform describes a store of metadata, knowledge, embeddings and memory, and names Notion and Confluence among its connector examples.

Source: DataHub Cloud context platform

Atlan lists knowledge management among its connector categories.

Source: Atlan connectors

The other scored vendors focus on structured metadata on the pages we read.

Unstructured sources are harder to govern than tables. A spreadsheet attached to an email has no owner field, no certification and no lineage. Documents mix confidential and public content in the same file. When an agent pulls a paragraph into its answer, there is often no column-level security to stop it.

That is why the vendors above present unstructured support as governance, not only as search. Collibra frames it within its control plane; Alation calls the feature Governed Collections. Ask how permissions on the source system are carried into what the agent receives, and what happens when a document's permissions change.

Source: Collibra homepage

Source: Alation press release: AIOS expansion with six products

For most data platform teams, tables first. The structured stack is where definitions, lineage and access rules already exist in machine-readable form, and where wrong answers do the most visible damage: a wrong revenue number travels further than a wrong paragraph. Once agents answer reliably on tables, documents can add the reasons behind the numbers.

There are exceptions. If your agents serve support or operations teams whose questions are mostly about procedures, documents may come first. In that case, weigh governance of unstructured sources heavily and test it with real permission boundaries.

Not as a separate criterion. It appears inside governance and integrations where vendors document it. If it matters to you, read the governance reasons on each review and the facts on the reviewed, not scored page, where Jedify's approach is described.

Lineage for documents is still thin across the vendors we read. A table has upstream sources and downstream dashboards; a policy document has authors, versions and readers, and its link to the data it describes is usually implicit. Some vendors connect the two through the glossary: a business term links to both the certified table and the document that defines the policy. When you evaluate unstructured support, ask whether the agent can follow that link, from a number in a table to the document that explains why it is calculated that way. If it cannot, the agent will have two separate sources of truth and no way to tell which one governs.

Take one folder with mixed permissions. Connect it, ask an agent questions as two users with different access, and check that each sees only what they should. Then revoke access to one document and repeat. Time how long the revocation takes to reach the agent. For tables, a delay of minutes may be acceptable; for a confidential document, it may not. Write the acceptable delay into your requirements before the proof of concept, so the result is a pass or a fail rather than an impression.