Week of Feb 4th

Enhancements

If your solution uses Generative AI, and you’re using the LLMs made available via LivePerson (not your own LLM), this announcement is for you.

LivePerson has made GPT-4o mini available to test and will automatically upgrade KnowledgeAI™ and the CoPilot Suite to GPT-4o mini on the following dates:

• North America: February 12, 2025 (11:30 PM EST)
• APAC: February 12, 2025 (3:30 AM AEDT)
• EMEA: March 30, 2025 

This upgrade will apply to solutions using the following Generative AI products: 

• KnowledgeAI™ 
• Copilot Suite:
  • Conversation Assist
  • AutoSummarization
  • Rewrite

Currently, these products use GPT-3.5 Turbo versions that are being retired from Azure OpenAI Service starting February 13, 2025.

LivePerson has developed new prompt templates optimized for GPT-4o mini based on our testing. If you're using the default, out-of-the-box system prompts, you can test the updates, but testing is not required.

Action Required:

• Test Your Prompt – If you are using a custom prompt, follow the instructions in our FAQs to evaluate how your custom prompt performs with GPT-4o mini. Please reach out to your account team with any questions.
• Report Any Issues – If you identify any issues before February 12, 2025, please report them HERE or contact your account team.

Our Commitment to You:

• If you report an issue, we will work with you to resolve it and provide an updated prompt that functions correctly with GPT-4o mini.
• You will receive an update on your reported issue every 48 hours until it is resolved.

Comparing GPT-4o mini to GPT-3.5 Turbo 

For several months now, LivePerson has been using GPT-4o mini in Early Access products not yet generally available. In short, we’re very pleased with GPT-4o mini’s performance. It’s just plain better: It responds to prompts with fewer hallucinations and is often faster.

Data inferencing by region

All data is processed within the appropriate region for GPT-4o mini. For instance, when using GPT-4o mini in EMEA, that data is processed in the UK. No data is stored per our agreement with Azure.

FAQs

1. When can I start testing 4o mini?

You can do this now unless your solution involves Conversation Assist. In that case, you can start testing on February 5, 2025. Refer to our earlier release note.

2. How do I test 4o mini with a LivePerson prompt?

1 - First, make a copy of a LivePerson prompt template.

2 - In Step 4 of the linked instruction in #1, you must select the LivePerson template that you want to copy. In the case of AutoSummarization and Rewrite, LivePerson has added new templates specifically for GPT-4o mini; copy one of these (1  below). In all other cases, LivePerson has updated our existing templates, so select any template that’s appropriate for your use case (2 below).

3 - In Step 6 of the linked instruction in #1, consider changing the prompt’s name, for clarity. 

Once you’ve completely created the copy, you can test it and evaluate the prompt’s instructions for potential changes. See our best practices. 

How do I test GPT-4o mini with my custom prompts?

1 - Open the Prompt Library.

2 - Make a copy of your custom prompt.

3 - In the copy, make the switch to GPT-4o mini on the Advanced tab of the prompt. This requires changing both the Subscription and LLM.

Tip: Consider changing the prompt’s name too, for clarity.

4 - Test the copy and evaluate the prompt’s instructions for potential changes. Here are direct links to prompt testing per product Summary, Rewrite, KnowledgeAI (for answer testing to support Conversation Assist and LivePerson Conversation Builder KnowledgeAI agents), and Prompt Management. See our best practices. If you’re interested in testing and need help, contact your LivePerson representative.

3. How can I tell if my model is hallucinating less?  

One of the benefits of adopting GPT-4o mini is that there are fewer hallucinations. 

When you move to GPT-4o mini, solutions using Retrieval Augmented Generation or RAG (for example, KnowledgeAI™ agents) might uncover cases where there is information missing in your knowledge base. When such content doesn’t exist and the model hallucinates less, you’re likely to see the model say, “I don’t know” more often. This is good, as it keeps responses grounded and surfaces knowledge/content gaps that you should address. 

4. Will Routing AI agents be updated, too? 

Routing AI agents currently use GPT-4, so LivePerson will update this solution at a later date.

For previous Release Notes on this topic, click here - 4o mini availability release notes 1/22

Enhancements

Reduce latency when processing consumer messages

A delay in the processing bot responses was removed when handling long bot messages, resulting in a slightly improved response latency performance

IBM Watson Connector: Support for Date response type

The IBM Watson connector will now recognize Watson responses of type "Date" and forward them to the LivePerson Messaging system as rich content of type Date Picker.

Fixes

Introduce separate rate limit when handling files as bot responses

When sending an array of files as a consumer to a bot connected via Third-Party Bots, a spam protection might have mistakenly triggered, discarding several files. With this release, file messages are not ignored when being sent as a list. 

Fix race condition when joining a conversation and send a message via public API at almost the same time

There was a race condition where the message event was received before the conversation change event in the joinConversationCommand, resulting in missed processing of conversation details. As a fallback, the worker fetched details from the Messaging History API, which was slow and not always reliable, potentially leaving conversationDetails empty.

The fix now ensures up to 3 retries to verify that the conversation is connected before resorting to the Messaging History API, leading to the correct order of handling events.

Features

WhatsApp Business Connector

WhatsApp Reply to a Message from Agent to Consumer

We are happy to announce that we now allow agents to reply to a specific consumer message on Agent Workspace to enhance communication by ensuring that your responses are clear and relevant!

Key Features:

1. Contextual Replies: When you reply to a message, the original message appears above your reply, providing context for everyone in the chat.

2. Button to Reply: Click on the reply icon (arrow) next to the message you want to respond to.

3. Multiple Responses: You can reply to multiple messages in a conversation, maintaining clarity in ongoing discussions.

Features

Features

WhatsApp Self Service for Applying a New Display Name

We are excited to announce that brands are now allowed to apply a business name change on WhatsApp channel within the Self Service via Channel Setup on the CCUI.

Updating the display name on WhatsApp is a two-step process:

1. In WhatsApp Business Manager:

• Log in to your WhatsApp Business Manager account.
• Select the relevant account and phone number.
• Navigate to Profile and click Edit Display Name.
• Once your name change request is approved, proceed to the next step.

2. In CCUI:

• Log in to CCUI.
• Go to Manage >Channel Setup > WhatsApp.
• Locate the relevant card and click Apply.

Limitations:

• For brands that do not own their WABA, please contact your LP representative. You can then proceed directly to step 2 in CCUI.
• The certificate for the new display name is only available for 14 days. If the certificate expires, you'll need to submit the display name for review again by Meta.
• If the user didn’t Apply the name change after 14 days > the Apply/Report button will disappear and the user will need to update the name again on their WABA.
• Brands can change their display names up to 10 times in a 30 day period.

Note: Please note that if your name change request has not been fully approved by Meta, you can click "Report Issue," and we will initiate a support ticket with Meta.