Prepare Reliable Knowledge for the Help Center AI Assistant - Article
Summary
An AI support answer can only be as useful as the knowledge that explains the real process. Clear prerequisites, current terminology and concrete outcomes help readers and the Help Center AI assistant identify the right guidance without inventing missing steps.
Maintain the source articles and their audiences first. Use synonyms, representative questions and administrative evidence to improve retrieval and diagnose issues, while keeping the Help Center assistant distinct from learning-content AI.
Use this article to:
- Prepare clear source articles and reusable FAQ answers
- Maintain terminology and test access-aware answers
- Use LLM usage and logs without confusing them with learning analytics
Know which sources the assistant uses
The Help Center assistant uses published Help Center articles, associated FAQs and release content available to the reader.
This assistant is separate from the LMS AI Learning Assistant. Courses and Resource Library documents follow their own retrieval and access model there; they do not automatically join the Help Center’s knowledge collection. Explain the distinction to users and content owners before deciding where information belongs.
The Introduction to the AI Learning Assistant owns the separate LMS learning-assistant overview; keep each assistant’s eligible sources and user purpose clear.
The Help Center assistant uses retrieval to find relevant source material and supply it as context for an answer. Maintaining an article is not the same as training a model on your organization’s knowledge. Grounding and response instructions are intended to reduce unsupported answers; they do not remove the need to verify important guidance.
Write sections that can answer a real question
Use descriptive headings and keep the explanation beside the control or decision it qualifies. Include what a setting changes, who can use it, prerequisites, the expected result and meaningful exceptions. Do not rely on a screenshot to carry a rule that is absent from the text.
| Weak source pattern | More useful source pattern |
|---|---|
| “Configure the options as appropriate” | Name the options, explain their effects and give a realistic reason to choose each |
| A heading followed by unexplained fields | Explain the task and relationship before a concise field reference |
| An instruction without a save or outcome | State what to save and how the reader can recognize the expected result |
| One long page mixing unrelated audiences | Use meaningful sections or separate canonical articles when the actions and permissions differ |
| A FAQ with no linked article | Attach the answer to the maintained article that owns the reader’s process |
Content is segmented around headings for retrieval and revised when the content is saved. Give each section enough context to be understood as a relevant answer, without repeating the complete article in every section. Search suggestions are generated from the content; they are not a manual list of approved answers.
For the current Help Center AI content pipeline, prepare English source guidance for reliable retrieval. The assistant can interpret questions in other languages and respond in the question’s language. This is different from translating the article itself or changing the Help Center interface language; agree additional source-language requirements before relying on a multilingual source collection.
Maintain useful synonyms
Open Settings → Synonyms as a Help Center administrator to maintain equivalent terms for the assistant. This is useful when users still search for a former product name or use a familiar business term for the same object.
Treat a synonym as an equivalence, not merely a related concept. “Participant” and “learner” may help describe the same person in your service; “certificate” and “course completion” can represent different outcomes and should not be treated as interchangeable without a valid process reason. After a terminology change, update the source explanation and test both old and new wording.
Test answers in the intended reader context
Use a small set of real support questions covering the main procedure, a prerequisite, an exception and a role boundary. Test as representative readers because a Help Center administrator’s broad access can conceal a missing audience selection.
| Scenario or check | Expected result or evidence |
|---|---|
| A direct how-to question is asked | The answer includes the applicable actions and links to the authoritative source |
| A synonym or former term is used | The assistant identifies the intended current concept without merging distinct objects |
| A linked FAQ answers the question | The response uses the eligible answer and points to its associated article |
| A reader lacks access to a source article | The restricted material is not disclosed as their answer |
| A required step is absent from all sources | The content owner adds the missing explanation rather than relying on the assistant to invent it |
| An article changes | Save the source, then recheck the affected question and source link |
Review usage and diagnose a failing request
Open Reports as a Help Center administrator. LLM usage provides aggregate requests and tokens; LLM log provides individual request evidence. These reports help distinguish source-quality issues from processing or capacity issues.
| Report or filter | How to use it |
|---|---|
| Client: All, API, System or User | Separate external API use, system indexing and suggestion work, and Help Center user conversations |
| Model type: All, Chat or Embeddings | Distinguish answer generation from embedding work |
| Year and month | Choose the period relevant to the issue or usage review |
| LLM usage totals and daily chart | Review request volume and input/output tokens; these are not a currency invoice or a learning completion measure |
| LLM log rows | Inspect date, client, IP address, model type, tokens and response information |
| A log entry’s Date | Open request details, including response code and request/response data, for focused diagnosis |
Logs can contain user questions, source context and other operational data. Limit access to people with a valid administration responsibility and share only the evidence necessary for the investigation. Clearing a reader’s visible chat does not erase these administrative records. Use the applicable service agreement for retention and data-handling requirements; the presence of these logs means the service should not be described as stateless.
If relevant source content is absent, correct the source or its access. If the log shows a processing or rate-limit issue, retain the time and response evidence and use the agreed support route. Provider and model limits depend on the deployed service; use the applicable configuration when investigating capacity.
Troubleshooting
| Problem | What to check |
|---|---|
| The answer omits a configuration dependency | Add the missing condition beside the applicable instruction in the authoritative source and retest |
| A FAQ is not used | Check the article association, publication state and reader access |
| Old terminology gives poor results | Update source wording and appropriate synonyms, then test both terms |
| An answer cites an unexpected process | Check ambiguous headings, overly broad synonyms and contradictory source guidance |
| A response fails after waiting | Inspect the relevant LLM log for processing or capacity evidence; retry as appropriate and retain useful details for support |
| A multilingual answer is unreliable | Check the English source guidance, the original question and the distinction between source language and answer language |
FAQ
-
Does saving Help Center content train the AI model?
The assistant uses retrieval to find eligible source material and provide context for an answer. Maintaining that knowledge is different from training the model itself.
-
Can the Help Center AI assistant use our Resource Library automatically?
Do not assume that. The Help Center assistant uses its own eligible articles, associated FAQs and release content, plus separately configured article integrations where enabled. The LMS AI Learning Assistant and Resource Library have a separate source model.
-
What is the difference between System and User in LLM usage?
System activity includes content indexing and suggestion generation. User activity represents Help Center conversations. API activity distinguishes external requests, such as an embedded website assistant.
-
Why should we avoid a synonym between completion and certification?
Synonyms express equivalence. Completion and certification can represent different rules and outcomes, so treating them as interchangeable can lead readers and retrieval toward the wrong guidance.