Two types: answer recommendations and bot recommendations

Two types of recommendations can be offered to your agents:

  • Answer recommendations: This is an article in a knowledge base that matches the query. The agent can send the answer to the consumer.
  • Bot recommendations: This is a bot that matches the query. The agent can delegate the conversation to the bot.

Two places: in the conversation versus in the widget

Recommendations can be offered in two places:

  • In the conversation window
  • In the On-Demand Recommendations widget

In-conversation recommendations are answers and bots offered automatically, in real time, and in line within the conversation with consumer. They're offered based on the consumer's query or message.

Animation showing the agent being offered an answer and a bot to respond to the consumer's query, where the agent selects the answer and sends it to the consumer

However, sometimes your agents need more flexibility when handling conversations. Sometimes, they need to be able to look up answers and bots on demand, regardless of what the consumer just said. On-demand recommendations offered in the On-Demand Recommendations widget meet this need. These recommendations are offered based on the agent's query, not the consumer's query or message.

Animation showing the agent searching on-demand for a recommendation to a particularly consumer query, where the agent selects the answer and sends it to the consumer

You can offer in-conversation recommendations, on-demand recommendations, or both.

Learn more about the agent experience.

Recommendations are intent-based

Conversation Assist offers recommendations based on the intent that’s detected. For an in-conversation recommendation, this is the intent detected in the consumer’s query. For an on-demand recommendation, this is the intent detected in the agent's query.

View of the agent workspace with two recommendations being offered to the agent

Recommendations are skill-based

Recommendations aren’t just intent-based, they’re skill-based too.

When Conversation Assist offers a recommendation, it always does so based on not just the detected intent, but also the conversation’s assigned skill. This means you can align your knowledge content and bots to desired skills.

You also have the option to further limit recommendations to agent groups and/or profiles.

When in-conversation recommendations are offered to an agent in a conversation assigned to skill A, if the conversation is subsequently transferred to skill B, those recommendations remain available to any agent that picks up that conversation on skill B. This is the case even though recommendations shouldn’t be offered to conversations assigned to skill B, i.e., no configuration rule for skill B exists in Conversation Assist. The recommendations remain available for 24 hours, as the cache of recommendations stored in memory is cleared automatically every 24 hours. We’re actively working on a fix for this issue, so stay tuned!

Answer recommendations: enriched via Generative AI

If you’re using Conversation Assist to offer answer recommendations to your agents, you can offer ones that are enriched by KnowledgeAI's LLM-powered answer enrichment service. The resulting answers, formulated via Generative AI, are accurate, contextually aware, and natural-sounding.

Answer recommendations: rich or plain

Currently, rich answer recommendations are supported only on the Web and Mobile SDK channels.

When offering your agents answer recommendations, you want them to be relevant. But you also want them to be engaging, right? We agree.

So, when it comes to offering answer recommendations, you have options: You can offer plain text answers. Or, you can offer both rich and plain answers, and let your agents choose which type to send within the conversation. Here below, we’ve done the latter.

Rich answers being offered to the agent along with plain answers, inline in a conversation

Rich answers being offered to the agent along with plain answers, via the On-Demand Recommendations widget

Considering supporting rich answers. Their multimedia nature makes them much more engaging than plain answers, leading to a best-in-class experience for the consumer. Learn more.

Process: how many recommendations to offer

You can specify the maximum number of recommendations (answers plus bots) offered by Conversation Assist at one time. You can offer up to 5. The default value is 3.

Max number of recommendations setting

Process: how recommendations are offered

Step 1: Retrieve answers

To begin, the system retrieves answers according to the rules specified for knowledge bases.

Answers are retrieved from a knowledge base using the KnowledgeAI search offering to match answers to the query. The number of answers that are returned is as follows:

  • If you aren't using an LLM to generate answers that are enriched via Generative AI, the system always returns one answer (article) from each knowledge base.
  • If you are using an LLM to generate enriched answers, the system sends the number of articles that you specify in the rule to the LLM for an enriched response. And it always returns one enriched answer per knowledge base.

Conversation Assist then filters the results to include only those that meet the min. confidence threshold specified in the knowledge base rule.

The remaining answers are added to a list of candidate recommendations.

Step 2: Retrieve bots

The system then retrieves bots via intent matching (pattern matching isn't supported) and according to the rules specified for bots.

Bots that match the consumer’s (or agent’s) query are retrieved as follows:

  • LivPerson Conversation Builder bots: These are retrieved by checking the intents in the bots’ dialog starters for matches to the query. Intent matches have scores that can and do vary. Conversation Assist filters the results to include only those that meet the min. confidence threshold specified in the bot rule.
  • Third-party bots: These are retrieved by checking for a bot that matches the query. Different providers provide different score ranges for intent matches, so Conversation Assist converts them to a universal score range. It then filters the results to include only those that meet the min. confidence threshold specified in the bot rule. Be aware that some third-party bots don’t provide intent match scores at all; in such cases, filtering to exclude any bots based on min. confidence is not applied.

The remaining bots are added to a list of candidate recommendations.

Step 3: Determine the recommendations to offer

The system evaluates the candidate lists of answers and bots and chooses those ranked highest by relevance score. The rules for this are as follows:

  1. Which recommendations are evaluated? First, include all answer recommendations. Second, include all bot recommendations. This means that answers are included before bots even when the answer scores are lower than that of the top bot recommendation.

  2. How are the recommendations ordered? Within each subgroup of recommendations (answers, bots), sort the recommendations by confidence score in descending order so that the higher the score, the higher the recommendation.

  3. Which recommendations to offer? Choose the recommendations to offer to the agent based on how the Maximum number of recommendations setting is configured.

For example, assume the system has compiled the following final candidate list, sorted below by confidence score:

  • Answer in knowledge base 1 = 90%
  • Answer in knowledge base 2 = 60%
  • Bot 1 = 100%
  • Bot 2 = 80%

If Maximum number of recommendations is set to 2, then only the answers are recommended to the agent. The agent doesn’t see any bot recommendations because answer recommendations are always included first. But if Maximum number of recommendations is set to 3, then the two answers and the top-scoring bot are recommended.