Internal Knowledge and RAG Assistants

The Problem: Finding the Right Answer, Not Just a Document

An employee looking for an expense policy, onboarding step or runbook may encounter many search results, an obsolete copy and no clear owner. Repeated questions then reach IT, HR or document owners even when approved guidance exists. A fluent assistant can make this worse if it supplies an uncited answer, quotes a document the employee cannot open, or treats an outdated exception as current policy.

The use case is a governed employee answer experience: find authorized, approved evidence quickly; expose the source and its status; and say when a reliable answer is unavailable. It is not a replacement for document ownership or a promise that every internal question can be answered.

From Employee Question to Verifiable Answer

Question → Identity → Authorized Retrieval → Hybrid Search → Rerank → Grounded Answer → Citation → Feedback. An employee asks in ordinary language. The system resolves the employee identity and entitlements before retrieving any material, then combines keyword matching and semantic search only over permitted content. Reranking selects relevant passages rather than treating the first match as authoritative.

The answer is drafted against the returned passages and links to documents the employee can actually open. If approved evidence is absent, contradictory or superseded without a clear successor, the assistant withholds a definitive answer and points the employee to the responsible owner. Feedback identifies a broken citation or missing document for human follow-up; it must not silently turn a bad answer into an unreviewed policy.

An Answer Pipeline with a Document Owner

For the underlying retrieval design, see the protected RAG solution; this page concerns the employee-facing business workflow, not a duplicate platform specification.

What Employees Can Ask About

Potential collections include SharePoint, Confluence, file shares, ticketing knowledge bases, wikis and policy repositories. Scope connectors by collection and owner; a source being technically searchable does not make it approved or universally visible.

• Policy assistant: explain the current travel or leave rule with its effective date, citation and owner for exceptions.
• Onboarding: where a new employee finds equipment, access and first-week instructions appropriate to their role.
• Technical documentation: an operator runbook or supported configuration step, with version and applicability stated.
• IT and HR FAQ: repetitive service questions whose answer has a maintained source, not informal ticket comments alone.
• Procedure search: locate the current approved step-by-step process and its applicability before an employee acts.

When a newer document supersedes an older one, exclude or label the old passage. If two current-looking sources conflict, surface that conflict to an owner instead of blending them into a confident answer.

Permission, Provenance and Refusal Controls

• Permission-aware retrieval: test source ACL mapping and revocation, not only login. A citation to an inaccessible page is both a broken experience and a possible exposure.
• Source ownership: record the steward, review status, version and effective date when present; define how retired documents leave the answer set.
• Citation verification: inspect whether each consequential claim is supported by the linked passage and whether the employee can follow that link.
• Refusal and escalation: abstain on inadequate approved evidence, unresolved conflicts and restricted content; route gaps to the document owner.
• Representative evaluation: test authorized, unauthorized, stale and conflicting questions; review retrieval, answer, citation and refusal separately.
• Redacted logs: keep the minimum query/answer trace needed for debugging without retaining unnecessary personal or confidential content.

What the Pilot Produces

• A collection and ownership register: source, steward, approval status, refresh path, ACL behavior and retirement rule.
• A retrieval/answer specification: identity boundary, indexed metadata, hybrid search, reranking, citation format and refusal behavior.
• A representative question set with expected source, access persona, acceptable answer or refusal, and reviewer decision.
• A content-gap queue for missing, ambiguous, obsolete and conflicting documents, assigned to the responsible owner.
• A redacted evaluation and operations dashboard design separating retrieval, answer grounding, citation validity and adoption.

These are proposed pilot artifacts, not an assertion that any customer repository has been connected.

Measure the Evidence Path Separately

Retrieval: for each test question, check whether the right current, authorized passage appears and whether denied material stays out. Generation: review answer correctness and appropriate refusal against approved evidence. Citation: verify claim-to-passage support, link accessibility and document version. Content health: count missing, stale and conflicting sources with an owner and resolution status. Operational use: track questions answered, escalations and repeat contacts without confusing volume with accuracy. Set targets only after a baseline and owner review; no outcome figures are implied.

Begin with One Owned Collection

1. Choose the boundary: one source collection with a named owner, known permissions and a clear current/retired status. Inventory documents and test accounts before embedding content.
2. Build the evaluation set: use representative real employee questions, redacted where necessary. Dozens, often roughly 50–100 when appropriate, can expose retrieval and refusal gaps, but no fixed count is a prerequisite. Include illustrative probes such as “Which leave policy applies today?”, “Can a contractor view the employee-only benefits rule?”, “Which onboarding checklist supersedes last year’s?”, and “Why do these two VPN runbooks disagree?”
3. Review before expansion: have source owners judge the retrieved passage, final answer, citation and refusal. Repair source hygiene and permissions first; add collections only after denial and freshness tests pass.

Risks and Refusal Tests

Related Architecture and Channels

The RAG solution explains retrieval architecture; data integration for RAG covers source connections, and private AI explores deployment boundaries. For spoken employee requests, see voice AI. Those technical approaches still require the document-owner, access and answer-quality decisions described here.

This is a design pattern, not a customer case study.

Frequently Asked Questions

Scope a Permission-Aware Knowledge Pilot

Bring one approved collection, its owner, representative questions and the access roles that should—and should not—see it. We can map the retrieval boundary, current-source checks and citation/refusal review before widening coverage.

Discuss the Pilot Boundary