Introduction
Developing an intelligent and helpful conversational AI solution hinges on the precise tuning of its core components: intents (consumer queries) and knowledge bases (answers). Intents allow the AI to understand what users want, while knowledge bases equip it with the information needed to respond effectively. This article highlights the crucial strategies for optimizing these elements, and takes you to relevant info that can help you to tune your Conversational AI solution.
Intent tuning
Carefully training intents is essential for building an effective and engaging user experience. Intents are the goals or actions that a user wants to accomplish. By giving the AI a wide variety of example phrases and situations for each intent, you help it to better understand the nuances of human language. This robust training enables the AI to accurately recognize what users are trying to say, respond appropriately, and ultimately deliver helpful, relevant, and satisfying interactions. Without well-trained intents, the AI may misunderstand user requests, leading to frustrating and unproductive conversations.
Refer to
Getting started with Intent Manager:
- What’s a domain?
- Add a domain
- Add intents to a domain
- Intent Manager best practices for intents, training phrases, and more
- Create a model version and release it
Automation-specific (Conversation Builder):
- Add intents or patterns to a dialog starter in a bot
- Specify patterns in interactions in a bot
- Specify Regular Expressions in interactions in a bot
Tuning-specific:
- Intent Manager: test a single utterance
- Intent Manager: test with the Model Tester
- Train and tune NLU
Knowledge base tuning
Just as well-trained intents enable understanding, thoroughly training the knowledge bases in your conversational AI application is important for delivering accurate, comprehensive, and helpful responses. The knowledge base serves as the AI's reservoir of information. By populating it with well-structured, relevant answers, you empower the AI to provide insightful and contextually appropriate info to your consumers. Conversely, a poorly trained knowledge base can lead to inaccurate, incomplete, or irrelevant answers, undermining consumer trust and the overall effectiveness of your solution. Investing in robust knowledge base training directly translates to more satisfying and valuable consumer interactions.
Refer to
Getting started with KnowledgeAI:
- What’s a knowledge base?
- Add a knowledge base: integrate a KMS/CMS or add one manually
- Add articles to a knowledge base: import content or add articles manually
- Best practices
- Search methods
Tuning-specific:
- Tune a knowledge base
- Optimize via dashboard analysis
- Evaluate consumer queries
- Resolve health issues in knowledge bases
Transcript review
Review of conversation transcripts offers invaluable insights into how your brand’s users interact with your conversational AI. By examining real-world dialogues, you can identify areas where the AI misunderstands user intent, provides irrelevant answers, or struggles with specific linguistic patterns. This analysis allows for targeted improvements to the AI's capabilities, ultimately leading to a more seamless, accurate, and satisfying consumer experience.
Which conversations to review?
Manually review at least 10 randomly selected conversation transcripts in each of the following categories:
- Conversations with low CSAT or MACS.
- Conversations where transfer to a human agent occurred.
- Conversations with containment, i.e., the conversation was resolved without transfer to a human agent.
- A random sample of conversations: select some conversations at random to uncover patterns that might not surface via the categories above; learn how “average” conversations went.
Which questions to ask?
As you review the transcripts, ask these questions:
- Understanding - Did the bot understand the consumer’s initial query (intent)? If No, what, if anything, could you change to improve the bot’s understanding?
- Did the flow of the conversation match the bot’s design, or are there errors? For example, is the welcome menu/message incorrectly surfacing too late in the conversation?
- Containment - Was the bot able to resolve the consumer’s query?
- Was the fallback flow triggered? If Yes, was the consumer’s message something that the bot should have understood?
More about conversations transferred to human agents
As you perform the transcript review, also ask these questions:
- Did the bot transfer the conversation at the right time? A bot only performs a transfer in certain scenarios, for example, when the fallback message is triggered multiple times, or at the end of specific dialog flows. Learn these by reviewing the design, and make sure the bot is only transferring the conversation when it is supposed to.
- How did the conversation go when the agent took over? Was the agent able to continue the conversation smoothly, did they repeat questions that the bot had asked? (Learn about automated conversation summaries, which can help greatly. Agent training can also aid here.) Did they understand what the consumer needed? Was the agent able to resolve the consumer’s query? If No, why not?
How to review transcripts
Access and use the Agent Workspace to manually review transcripts:
- Click All Conversations.
- Click the Filter icon to filter the conversations according to specific criteria.
Use the Conversation Status filter to view only closed, open, and overdue conversations. Use the Skills filter to view only conversations with specific bots.