Configure an existing product
- The source set and audience are bounded.
- Supported connectors preserve the required access model.
- Read-only answers with citations solve the actual job.
- The built-in administration and logging meet the review need.
Give the team internal answers they can inspect, not just accept.
We turn approved company records into a permission-aware knowledge system with visible evidence, freshness and conflict rules, a real evaluation set, and a deliberate boundary around action.
The implementation boundary
Retrieval quality begins before the vector database.
A smooth answer can still be wrong for the business if the system retrieves a superseded policy, a draft contract, another client's record, or two sources with unresolved authority. The implementation makes those operating rules part of the product.
- Authority belongs to a fact and process, not a software brand.
- Permissions must survive the full route from user identity to retrieved passage.
- Freshness and conflict need explicit behavior.
- Retrieved content is treated as untrusted input, not system instruction.
- Write access arrives only after read-only assurance passes.
- Question and fact map
Representative employee questions, expected answers, required facts, decision owners, and the work each answer supports.
- Source authority register
Primary and supporting records, content owners, scope conditions, review dates, and conflict-resolution rules.
- Access and retrieval design
Identity path, permission filters, tenant or department boundaries, indexing choices, and citation behavior.
- Evaluation set
Known answers plus restricted, missing, stale, contradictory, and hostile-content cases.
- Implementation and handoff
A bounded working route, operating responsibilities, review cadence, incident evidence, and a decision on future actions.
Build versus configure
Custom work has to buy a specific operating capability.
If an existing product can answer from approved sources under the required permissions, use it. Custom retrieval is valuable only when the business can name the boundary it needs to control.
Consider custom implementation
- Evidence must be assembled across unsupported or operational systems.
- Client, role, row, or field boundaries need application-specific enforcement.
- Source authority and conflict rules differ by process.
- The answer must enter a controlled approval and action route.
Implementation questions
What to settle before connecting the company knowledge.
The hard part is deciding what the organization considers true, current, and permitted.
Is an AI knowledge base the same as a Business Brain?
No. An AI knowledge base retrieves and summarizes approved material. A Business Brain adds decision context, approval, controlled action, and outcome evidence when the workflow requires them. Many teams should begin with read-only knowledge retrieval.
Can we use our existing documents and systems?
Usually. The implementation maps authoritative facts to their current systems, then evaluates the connectors, identity model, metadata, update behavior, and retrieval quality. Connecting a folder is not enough if ownership and access rules are unresolved.
Will every answer include citations?
The target is to show the evidence supporting material claims and make missing or contradictory evidence visible. A citation proves where a passage came from; it does not by itself prove that the source is current or authoritative, so those rules are part of the implementation.
Do we need a custom RAG platform?
Not automatically. Built-in company-knowledge features or an existing knowledge system may satisfy the first use case. Custom retrieval is justified when unsupported sources, tenant boundaries, authority rules, evaluation, or controlled workflow needs exceed those products.
When can the knowledge system take action?
Action is considered after read-only answers pass permission, freshness, conflict, and hostile-content tests. Every write tool then needs a narrow business purpose, validated inputs, an approval rule where required, and an outcome record.