When a consumer asks a question or makes an intentful statement, and you pass that query to a knowledge base to find an answer, the answer can be retrieved in one of two ways:

  • AI Search
  • Intent Match

Neither search method makes use of Generative AI or an LLM, but you can do that too - to enrich the answer and make it more contextually aware and natural sounding. Learn more here.

AI Search is KnowledgeAI™’s powerful, one-size-fits-all search method that’s based on the latest research in deep learning. Key characteristics of this search include the following:

  • There’s no setup required: This is some of the best news. AI Search works right out of the box. Don’t want to invest the time in creating and tuning intents in Intent Manager? No problem. You don’t have to. Just add your content to the knowledge base. Then expose it via a bot integration (use the KnowledgeAI interaction), or set up the knowledge base for Conversation Assist as normal. That’s it.
  • It gets to the intent: The AI Search isn’t a type of search that just looks at the words for matches found between the consumer query and the answer. The search also uses language embedding models to make a best attempt to determine the meaning behind those words: What is the consumer trying to convey?
  • It’s context-aware: The search considers the whole of the consumer query (and nothing more at this time). So, for example, “I want to book a flight” is understood differently from “Can I borrow a second book at the library?” Homonyms like these are handled.
  • It’s phrasing-agnostic: Synonyms are handled too: “I want to buy a ticket” and “I want to purchase a ticket” are understood similarly. Keep in mind that certain niche or archaic synonyms, especially abbreviations, might not be supported (unless you add them directly to the articles).
  • It returns multiple answers: If more than one good answer is found, multiple answers can be returned. This means you can use the KnowledgeAI interaction to populate a “carousel” of good answers to the consumer, providing a better consumer experience.

An Intent Match search is one that makes use of Natural Language Understanding or NLU to find the right answer to the consumer’s query. NLU leverages AI to identify various attributes of a message: meaning, sentiment, and more.

To use this type of search, you must add a set of intents in Intent Manager and tune them for optimal answer retrieval. Then, back in KnowledgeAI, you must tie your articles to the intents.

Like the AI Search, the Intent Match search is powerful and effective.

Choosing a search method

One common question we get is, “Which search method do I use: Intent Match or AI Search?” This is a good question, and our answer to this is to use them both if you can, and if it meets your needs. They are better together.

  • Works right out of the box. No setup is required.
  • Performs well when handling consumer queries that are similar to articles in knowledge bases. The broader the coverage, the higher the probability that there will be a match to the consumer query.
  • Makes up for issues that can occur when using intents. Since intents are trained by humans, an intent model might have some flaws: There might be gaps in the coverage. There might be an overlap in the coverage. Or, an intent might not be trained well enough. In other words, there’s an article that could be returned, but the match to the input query doesn’t meet the threshold required to return the article. Using the AI Search as a fallback to the Intent Match search is a good, recommended strategy to address issues like these.
  • Lets you manipulate outcomes by tuning the answer retrieval yourself. This gives you a layer of control.
  • Can outperform AI Search. NLU is better at making predictions. So, if you take the time to tune an intent well, the intent will likely outperform the AI Search. (That said, there’s a tradeoff here, as it’s a fairly significant effort to add and tune intents. Keep in mind that the AI Search performs well on its own.)
  • More suitable for specialized scenarios. Conversations that use a highly specific vocabulary do better when intents are used.
  • Sometimes, only an Intent Match does the job. Scenarios can exist where the consumer query properly matches to an article that doesn’t have any content that relates to the query.

Search offerings

When you select an answer retrieval method in various parts of the UI, the following choices are offered:

  • KnowledgeAI (recommended)
  • Intent match only
  • AI search only

Here’s an example:

The list of search offerings in the Answer Tester

The “Intent match only” and “AI search only” options are primarily intended for testing and for diagnosing issues during troubleshooting. In Production, we recommend that you use the “KnowledgeAI” offering.

“KnowledgeAI” option

The “KnowledgeAI” option is what’s used by the KnowledgeAI interaction in LivePerson Conversation Builder.

  1. Runs an exact match search.
  2. Runs the Intent Match search.
  3. Runs the AI Search.

“Intent match only” option

  1. Runs an exact match search.
  2. Runs the Intent Match search.

“AI search only” option

  1. Runs an exact match search.
  2. Runs the AI Search.

Search flow

Whether searching one knowledge base or several, the flow is the same:

  1. Using the specified search offering ("KnowledgeAI," "Intent match only," or "AI search only"), find the article matches in the knowledge bases.
  2. Aggregate the results, and keep only those in one of the following groups, listed in order of priority:

    1. Exact match
    2. Intent match
    3. AI Search match
  3. Sort the results in the group by match score.
  4. Return the top results based on the number of answers requested.

The search flow in KnowledgeAI, which in order checks for an exact match, an intent match, and finally an AI Search match. If an answer is found, it is optionally sent to the LLM service for enrichment, and the result is returned. If no answer is found, no result is returned.

Match scores and thresholds

Match scores

Whenever a knowledge base is searched for an article match to a consumer query, the results receive scores. The score indicates the level of confidence in the match: VERY GOOD, GOOD, and so on.

AI Search assigns scores as follows:

  • VERY GOOD: 85-100% match
  • GOOD: 70-85% match
  • FAIR PLUS: 65-70% match
  • FAIR: 50-65% match
  • POOR: 0-50% match

Intent Match search assigns scores using the scoring breakdown for the NLU engine used by the associated domain.


Other applications in the Conversational AI suite let you specify a threshold — that is, a minimum score — that a result must have for it to be returned as an answer.

  • You might be sending answers to consumers in automated conversations with bots. In this case, you can specify a threshold in the KnowledgeAI interaction within the bot in Conversation Builder.
  • Or, you might be recommending answers to agents inline in their conversations with consumers. And, in this case, you can likewise specify an answer threshold in Conversation Assist’s settings.

Regardless of the integration, if you lower the threshold, be sure to test whether the quality of the results meets your expectations. It's generally recommended to keep the quality above FAIR PLUS.


Which search method is best depends on the use case, your requirements, and your resources. KnowledgeAI’s AI Search works very well on its own and requires no setup. But use them both if you can; they are better together.

AI Search isn’t performing as well as I’d like. What can I do?

If you have a consumer query for which there isn’t a relevant article to serve as the answer, just add that article.

While it’s unlikely that an existing, relevant article won’t yield results, it might happen. In this case, improve the article’s title and/or add tags to the article.

Is AI Search available for external knowledge bases?

AI Search isn't available for external knowledge bases that don't use LivePerson AI. Knowledge bases of this type use the CMS' query and answer API for article suggestions/answers.

How does AI Search work?

In the internal knowledge base, AI Search performs 3 steps to get results:

  1. Pre-processing: This includes a simple query classification process to remove small talk (chitchat) from the input query. It also includes splitting a long query with multiple sentences into single sentences.
  2. Retrieval: The input query is concurrently checked against:

    • Title alone
    • Summary, Detail, and any tags
    • Title, Category, and any tags
  3. Post-processing: This includes the following:

    1. De-duplication: From the retrieved articles, duplicate articles or highly similar articles are removed.
    2. Re-ranking: The articles are reordered based on score and category.
    3. Score calibration: The internal scores of the articles are calibrated to VERY GOOD, GOOD, FAIR PLUS, etc.

How does the Intent Match search work?

In the internal knowledge base, the input query is checked against:

  • title
  • domain intents

How does the exact match search work?

In the internal knowledge base, the input query is checked against:

  • title