Go to Kayako Classic
Contents
Overview
An eligible, published Help Center article is surfaced and linked by the ATLAS AI assistant for some requests but not for others. In one interaction, ATLAS returns a "no answer found" result and hands the request off to a human; in another interaction — with a more simply worded request — ATLAS correctly finds and links the same article.
Symptom: For a simple, single-topic request (for example, "W-9 Request") ATLAS returns the correct article automatically. For a similar request whose subject mixes multiple topics (for example, "Gift Matching, w-9 needed"), ATLAS reports low confidence / "no answer found" — even though the article exists, is published, and is publicly accessible.
Note: This is not a defect, and it is not an article eligibility or indexing-delay problem when the article already meets all standard requirements. The inconsistency is driven by how ATLAS extracts search keywords from the request, combined with the target article having very little indexable body text. The resolution requires no code or configuration changes.
Prerequisites
Before you begin, confirm the following:
- You have an administrator or knowledge base editor account with permission to edit Help Center articles.
- The affected article already meets ATLAS eligibility requirements. If it does not — for example, if it was only recently published — resolve that first. For eligibility and indexing-cadence guidance, see New Knowledge Base Content Not Appearing Immediately in the Kayako ATLAS Plugin.
Why This Happens
ATLAS answers a request in two stages, and the failure occurs in the first stage:
- Keyword extraction: An AI step reads the ticket subject and first message and extracts the most relevant keywords — error codes, technical terms, functionalities, or the type of issue. These keywords drive the search.
- Knowledge base search: The extracted keywords run a similarity (vector) search against your articles. ATLAS only uses an article if it ranks among the top results with a strong enough similarity score. Articles that do not rank high enough never enter the AI's context.
When the subject line combines multiple topics, the keyword extraction can lean toward the wrong topic and pull the search away from the intended article. Compare two requests for the same document:
| Subject | Extracted intent | Outcome |
|---|---|---|
| "W-9 Request" | Single, direct — maps cleanly to the W-9 article | Article retrieved; ATLAS replies with the link ✅ |
| "Gift Matching, w-9 needed" | Mixed — the extractor treats "gift matching" as the primary topic | Search drifts away from "W-9"; article not retrieved ❌ |
When the search drifts toward the secondary topic, the target article can only remain in the results if it has enough matchable content. If the article body is minimal — for example, a single line such as Find a copy of our W-9 attached. with the actual document provided only as an attachment — it has almost no indexable text. It cannot score in the top results with a high enough similarity, so it drops out entirely. Because it never enters the AI's context, ATLAS reports "no answer found."
Note: Attachments and empty or external anchor links provide little to no indexable text for the search. The searchable content is the article's body text.
Solution
Enrich the affected article's body text with more descriptive, keyword-rich content that covers the varied ways people phrase the request. This raises the article's similarity score across different query intents so it stays in the retrieved results even when the subject mixes topics. No code changes are required.
1. Identify the affected article
- Locate the article that ATLAS fails to surface (for example, at
https://your_instance.domain.com/article/<article_id>). - Note whether its body is minimal, or whether the actual content lives only in an attachment or an external link.
2. Confirm eligibility is already met
This confirms that content — not configuration — is the issue:
- The article status is Published.
- The article loads in a private/incognito browser without signing in.
- The article's parent section/category is set to Public.
- The article was last updated beyond the observed indexing cadence (approximately every 4 hours).
3. Add descriptive body text
- Replace or supplement the minimal text with content that names the document and describes its common use cases and synonyms. For example:
"Our W-9 tax form is available for download here. This document is commonly requested for gift matching programs, corporate matching gifts, nonprofit/donor verification, and donation tax documentation purposes."
- Include the terms people actually use when asking — both the direct name of the document and the related workflows or contexts in which it is requested.
4. Keep attachments, but don't rely on them for matching
- If the real document is an attachment, keep it — but make sure the searchable body text describes it fully, since attachments and empty anchor links provide little to no indexable content.
5. Save and allow re-indexing
- After saving, allow time for ATLAS to re-index. Indexing has been observed on an approximately 4-hour cadence, though this is an observed pattern rather than a guaranteed interval.
6. Verify the fix
- Submit test requests using different phrasings, including a mixed-topic subject that previously failed (one that leads with a secondary topic and mentions the document only as a detail).
- Confirm ATLAS now surfaces and links the correct article for each phrasing.
- Compare against a simple, single-intent request to confirm consistent behavior across both.
Key takeaway: When an eligible article is only inconsistently surfaced by ATLAS, the most common cause is thin article content combined with mixed-topic request wording. Enriching the article body with descriptive, synonym-rich text is usually enough to make retrieval reliable — no code or configuration changes needed.
Frequently Asked Questions
- 1. How do I know this is my issue and not a standard eligibility or indexing problem?
- If your article is Published, loads in an incognito browser without login, sits under a Public section, and was last updated well beyond the ~4-hour indexing window — yet ATLAS still surfaces it for some requests but not others — then eligibility is not the cause. Inconsistent behavior across differently worded requests for the same article points to keyword extraction combined with thin article content.
- 2. Why does ATLAS answer some requests correctly but fail on similar ones?
- ATLAS extracts keywords from the ticket subject and first message to drive its search. A simple, single-intent subject (for example, "W-9 Request") matches your article cleanly. A subject that mixes topics (for example, "Gift Matching, w-9 needed") can lead the extractor to focus on the secondary topic, pulling the search away from the intended article. If the article has little body text, it drops out of the search results entirely.
- 3. My article's real content is an attachment or an external link. Does that matter?
- Yes. Attachments and empty or external links provide little to no indexable text for the search. Even if the document is attached, add descriptive body text that names the document and describes its common uses and synonyms, so the article can be matched reliably.
- 4. What exactly should I add to the article body?
- Add a concise description that includes the document name plus the varied contexts and phrasings people use when requesting it. For example: "Our W-9 tax form is available for download here. This document is commonly requested for gift matching programs, corporate matching gifts, nonprofit/donor verification, and donation tax documentation purposes."
- 5. Do I need a code change or engineering intervention to fix this?
- No. This is resolved entirely by enriching the article's body content on your side. No code or configuration changes are required.
- 6. I updated the article but ATLAS still isn't surfacing it. What should I check?
- Allow time for re-indexing (approximately every 4 hours) after saving, then re-test with multiple phrasings, including the mixed-topic wording that previously failed. Confirm the descriptive text is in the article body itself (not only in an attachment or external link) and that the article still meets all eligibility requirements. If it still fails consistently after that, contact Support and include the specific request wording that isn't working, the article URL, the article's publish status, confirmation that it is publicly accessible and under a Public section, a sample request that should return the article, and the actual ATLAS response received.
ATLAS KBA
Comments