OpenMaxUse Case Guide
Defined
approved source scope
Required
passage-level citations
Matched
user and document access
Visible
freshness and uncertainty
TL;DR
- Problem: Employees cannot reliably find the current answer when knowledge is scattered, duplicated, stale, or permissioned differently across systems.
- Approach: A RAG workflow retrieves relevant approved passages first, then uses them to prepare an answer with citations and an uncertainty signal.
- Control: The knowledge system remains authoritative; the agent answers, identifies gaps, and hands unresolved questions to the source owner.
What is an AI knowledge base agent?
An AI knowledge base agent combines retrieval with a language model so answers are prepared from approved company content instead of model memory alone. A production design must preserve source ownership, document permissions, version and freshness information, citations, review, and a path for unanswered questions.
When a knowledge base agent is useful
| Situation | What the agent can do | What stays with people or systems |
|---|---|---|
| Internal policy questions | Retrieve the current approved policy and cite the relevant passage. | Policy ownership, interpretation of exceptions, and final approval. |
| Support knowledge | Prepare an answer from procedures, product notes, and resolved cases within the user's access. | Account actions, policy exceptions, and sensitive customer decisions. |
| Onboarding and operations | Answer process questions, link required forms, and identify missing or conflicting guidance. | Access grants, HR decisions, and source-document maintenance. |
A governed RAG workflow in four stages
1
Approve and inventory sources
Record the source owner, audience, permission model, version, review date, expiry rule, and excluded content.
2
Index with permission metadata
Preserve document identity, section location, access groups, modification time, and deletion state during ingestion.
3
Retrieve and answer with citations
Retrieve only content the requester may access, show the supporting passage, and state when evidence is incomplete or conflicting.
4
Review gaps and maintain sources
Send unanswered, low-confidence, stale, and disputed questions to the correct owner, then update the source rather than patching only the answer.
Quality and security checks
| Check | Acceptance question | Evidence to retain |
|---|---|---|
| Retrieval | Does the result include the correct document and passage for representative questions? | Question, retrieved passages, ranking, reviewer judgment, and missed source. |
| Answer grounding | Can each material statement be supported by the cited passage without adding unsupported detail? | Answer, citations, unsupported statements, corrections, and reviewer decision. |
| Permissions | Can users retrieve only documents they are already authorized to view? | User identity, access group, source permission, denied request, and permission change. |
| Freshness | Do updates, deletions, replacements, and expired documents change answers within the agreed window? | Source version, index time, answer version, stale-result incident, and resolution. |
How to measure a pilot
Build a representative question set by team, source, permission level, and difficulty. Include normal questions, ambiguous wording, stale documents, conflicting policies, restricted files, and questions with no approved answer.
| Metric | What it reveals |
|---|---|
| Retrieval success | Whether the correct approved passage appears in the reviewed result set. |
| Grounded answer acceptance | Whether reviewers accept the answer and citations without material correction. |
| Permission accuracy | Whether restricted content is consistently excluded for unauthorized users. |
| Unanswered-question quality | Whether the agent refuses safely and routes the gap to the correct source owner. |
Pilot gate: Expand only after retrieval quality, citation support, permission enforcement, freshness, refusal behavior, and owner response meet the thresholds agreed for each knowledge domain.
Frequently asked questions
How is a knowledge base agent different from a document chatbot?
A document chatbot can be useful for a small, static file set. An enterprise knowledge base agent also needs continuous source updates, permission-aware retrieval, citations, audit records, owner handoff, and controls for actions that follow an answer.
Can a RAG workflow eliminate unsupported answers?
No. Retrieval can reduce unsupported answers, but poor indexing, wrong permissions, stale documents, weak prompts, or conflicting sources can still produce errors. Use citations, confidence rules, refusal behavior, and human review for important decisions.
How should private knowledge be deployed?
Choose the deployment path from data classification, residency, identity, network, retention, and vendor-access requirements. Verify where documents, indexes, retrieved passages, prompts, logs, and backups are stored before connecting sensitive sources.
Build a knowledge workflow your team can verify
Start with one owned knowledge domain, a representative question set, and clear acceptance rules.
Explore OpenMax Agent Cloud