On this page

Best practices for documentation that supports AI

Turing works best when articles and QnAs in the knowledge base adhere to certain guidelines. The old computing adage of “garbage in, garbage out” applies to AI as much as to earlier technologies. Most of these guidelines are typical for professional/technical writing, especially content that has requirements for accessibility and localization.

To enable searching through the knowledge base, Turing cuts up articles into smaller chunks (paragraphs and sentences). As an overarching principle, think of the various elements of an article or QnA as modules that may be used in a variety of ways.

Multiple mediums/modalities

  • Include supporting text clearly explaining images and videos. Every image and video should have alt text at minimum for a variety of reasons; lengthy discussion should be in the body of the text. Images and videos (for now), are not used to train the model.
  • Rather than include pictures of text, include the text itself in the body.
  • Minimize the use of tables. Two-column tables can easily be represented with a description list (HTML dl element) or nested ordered/unordered lists.


  • Avoid long articles. Ensure tight cohesion within an article, meaning that it covers only a single topic.
  • Ensure that the question in a QnA pair is properly-phrased as a question, and that the answer directly addresses to the type of question (for example, yes or no for a yes/no question, a procedure for a how question, an explanation for a what question).
  • Include URLs of relevant articles in QnAs.
  • Judiciously use pronouns (it, that), and ensure that the antecedent for pronouns is clear. Prefer using the name of the object, product, or feature.


  • Ensure tight cohesion within paragraphs.
  • Dedicate a paragraph in articles to cover common questions.
  • Each paragraph should make sense when read in isolation, without the paragraphs or sections before or after it. Avoid using transition words or phrases—such as in addition, next, or also—to start paragraphs.


  • Write your documentation in the same voice you would want your AI agent to talk to your customers.
  • Consistently use approved product terminology. Avoid colloquial or internal terms and phrases, including jargon and code names.