What is an enterprise knowledge base? Establish sources, versions and access

An enterprise knowledge base contains knowledge an organization organizes, maintains and makes available for use. It can reside in a wiki, document-management system, shared folder or other tool. Generative AI is not a prerequisite for being a knowledge base.

Adding search or AI answers requires assessing how an application obtains applicable material, enforces access, handles changes and checks results. Existing systems can retain formal records. Design additional capabilities for the actual problem rather than declaring those systems unusable.

Identify the user's task first

Users may need to find a document, compare versions or prepare a source-linked answer. Different tasks can use different tools and processing methods.

Consider a hypothetical manufacturing SOP query: one procedure is approved while another remains under revision. Establish which applies to current work, who may use it and who resolves a conflict. Searchability alone cannot establish authority from a filename. This is a planning example, not a client Q&A result or effort record.

Six source-readiness questions

Metadata can support source, version and access rules. A label is not proof of correctness or approval. Ownership and application behavior still need definition.

Document systems and wikis can provide sources

Atlassian's Rovo Search documentation describes searching Confluence, Jira and connected sources, with source-supported answers. It provides a concrete example of wiki content supporting an AI-search product, not evidence that every unconfigured folder already has those capabilities.

Assess existing systems and interfaces before choosing connectors, parsing, indexing, search or other processing. Microsoft's RAG overview discusses different source and retrieval designs. A language model and an application's vector-search pipeline are not the same component.

Test formats against the chosen processing method

Scans and diagrams may require parsing or OCR, particularly for a selected text-search pipeline. Models are not universally restricted to previously extracted plain text. Anthropic's PDF support processes PDF text and page images directly, subject to platform, request, context and quality limits.

Page count alone cannot establish that a document is unusable. Accepting a file cannot establish correct reading either. Test table relationships, footnotes, legends, scan quality and conditions spanning pages with actual formats. Unrecorded oral knowledge also needs authorized capture and confirmation before an application can use it.

Make updates and authorization checkable

Agree authoritative versions, drafts, effective dates, conflicts and withdrawal with source owners. After an update, confirm relevant processing succeeded and test old versions, deletion and access changes. Microsoft's document-access guidance requires identity, document permissions and query enforcement in its product. Connecting a model does not automatically provide authorization.

For generated answers, assess whether sources support statements, limits were omitted and missing evidence is handled. Citations make checking easier; they cannot replace checking.

Choose the next step from the gap

Where approved versions or owners are missing, resolve those prerequisites. Where usable sources exist but documents are hard to find, assess existing search. For cross-source preparation and generation, consider RAG's uses and boundaries.

The existing RAG service can help discuss scope. Representative documents, common questions, access requirements and owners clarify what the first phase needs to establish more directly than selecting a model first.