You are here:
How Does Einstein Article Recommendations Work?
Einstein Article Recommendations helps support service reps resolve customer cases efficiently by recommending knowledge articles that were attached to similar cases in the past. Reps don’t have to waste time searching or scrolling through lists of articles, and can quickly attach recommended articles or dismiss them as not helpful.
Required Editions
| View supported editions. |
Article recommendations are generated by an AI model that continually retrains itself and improves its predictions as the underlying set of case and article data grows. The model is trained using various natural language processing (NLP) features that capture overlaps between case and article data. The model’s algorithms rank candidate articles based on their relevance to an open case, and only articles that are deemed highly relevant are recommended.
Determining Article Relevance
To determine an article’s relevance to an open case, Einstein Article Recommendations considers several factors. These factors include:
- Language: The language that Einstein detects in the case and the article’s language
- Term overlap: How closely key words in the case match text in the article
- Attachments: How often service reps attached the article to similar cases
- Dismissals: How often service reps dismissed the article as not helpful on similar cases
- Term span: The distance between certain case terms in the article
- Longest common subsequence: The length of the longest common text sequence between the case and article
You customize your model by choosing the case and knowledge article fields and the languages for Einstein to consider.
| Object | Fields | Details |
|---|---|---|
| Case |
|
The Subject field is usually populated and summarizes the issue, while the Description field provides details. If you’re using custom fields in place of Subject and Description, you can configure Einstein Article Recommendations to refer to them instead. |
| CaseArticle |
|
The CaseArticle object represents the attachment of an article to a case, and provides key data for article recommendations. |
| KnowledgeArticleVersion |
|
The KnowledgeArticleVersion object represents a specific version of an article. Einstein Article Recommendations only recommends the published versions of articles. Because this object doesn’t include a body field, the custom text fields on article versions are combined into a custom “Article Body” field for analysis. Einstein also considers the Language field on this object when making recommendations. |
Language Considerations
Einstein Article Recommendations is available in Dutch, English, French, German, Italian, Portuguese, and Spanish. Einstein uses a single model to generate article recommendations in these languages.
In a new org, you select which languages and fields for Einstein to use in the predictive model. If you’re already using Einstein Article Recommendations, we preselect languages for you. If you modify the selected languages in Setup, rebuild your model to show recommendations in those languages. Make sure that the selected languages are active in your Knowledge settings. You can add articles in a supported language without rebuilding the model.
Einstein uses a language prediction model to detect the language on a case. The Language field on a knowledge article version identifies its language.
- If the case language is supported, Einstein recommends relevant articles in that language. If there aren’t any relevant articles in the language, then no article recommendations appear.
- If the case language isn’t supported, Einstein recommends relevant articles in the org’s Knowledge primary language.
- If both the case and Knowledge primary languages aren’t supported and English is a selected language, Einstein recommends relevant English articles.
These considerations mean that a case and its recommended articles can use two different languages.
Feature Design
Here’s what takes place behind the scenes, from case creation or update to article recommendation.
Step 1: Case Creation or Update. A case is created, or a field included in your article recommendation model is updated on a case.
Step 2: Search. Einstein Article Recommendations intercepts the case and conducts multiple searches using the case’s language and terms from case fields in your model. The search results are combined into the candidate article set, which can include new articles with no previous attaches.
Step 3: Re-Ranking. Articles are re-ranked. A set of NLP features is computed for each case-article pair. The model evaluates each article’s feature set and uses prediction probabilities to rank the articles for recommendation.
Step 4: Recommendation. Articles above the recommendation threshold are classified as positive and recommended to the service rep on the case.
Step 5: Service Rep Interaction. The service rep interacts with the recommendations by clicking, hovering over, accepting, or dismissing them. Their actions are recorded to improve future recommendations.
Step 6: Retraining. Your model is periodically retrained, or rebuilt, to capture changes to your case and article data. Relevant changes include new cases, new articles and article versions, and new case-article attaches. The retrained model replaces your current model only if Einstein determines that the retrained model will provide better article recommendations. Retraining completes within a day, and the new model is put to work immediately. This automatic retraining isn’t visible in the UI and doesn’t require action from the admin.
Einstein Article Recommendations uses Amazon Web Services (AWS) as a third-party hosting provider. For details about infrastructure and sub-processors, see Trust and Compliance Documentation.
What’s the difference between Einstein Article Recommendations and Suggested Articles?
Like Einstein Article Recommendations, the Suggested Articles feature suggests knowledge articles in the Lightning Service Console. However, it relies only on keyword-based search and can’t refine its suggestions or incorporate data from past cases.
For example, many cases contain the same subject, description, and category. Because Einstein Article Recommendations considers which articles were attached to similar cases in the past, the top recommended article is likely to address the issue. Suggested Articles can search your knowledge base for articles containing case keywords, but the exclusion of case data and AI features means that service reps can waste time searching through article suggestions to find the one they need.
If you use Einstein Article Recommendations, we recommend disabling Suggested Articles to ensure a clean user experience for your support team. Otherwise, service reps see two sets of articles that relate to the case.
