Using the AI chat
Interact with AI-powered assistants, AI agents, and skills by using the AI chat. With this chat, you can complete various generative artificial intelligence (AI) tasks, or even use AI assistants to perform more complex tasks like sending emails or posting job requisitions.
To access the AI chat, click Chat in the menu. It opens a chat for you to start the conversation, click New chat + on the side panel to start a new chat.
Using AI assistants in the chat
When you first interact with the AI agent in the chat, it can manage general AI tasks. However, when your tasks become more focused on a specific domain, you can smoothly route the conversation to AI assistants designed for a domain or use case.
Using AI assistants in the AI chat requires that you complete the following steps beforehand:
- Create an AI assistant.
- If you want to use skill-based actions, you must connect to the apps that you want to use the skills.
- Create actions in that AI assistant.
- Publish the contents of the AI assistant to the draft and live environments.
- Add the AI assistant to the chat.
If you don't add AI assistants to the chat or your prompts don’t match any AI assistant added, the conversation is handled by the AI agent and its base LLM.
Interacting with the AI chat
You can interact with the AI chat to assist you with your tasks by:
- Prompting questions and simple tasks to the AI agent to generate content for you.
- Use a specific AI assistant by mentioning it.
- Expecting the AI agent to route the conversation to an AI assistant based on the conversation context.
Prompting questions and simple tasks to the AI agent
The AI agent uses large language models (LLM) to create general-purpose answers and assist with a wide range of generative AI tasks, such as question-and-answer, summarization, classification, text generation, information extraction, and translation for various topics. This toggle is disabled by default.
Explore the sample prompts to get started with the AI agent.
Use prompt engineering techniques to ensure that the AI agent can accurately understand and respond to your prompts. For more information about how to write prompts for AI agents, see Tips for writing foundation model prompts: prompt engineering in the IBM watsonx documentation.
When you send a prompt to an AI agent, its LLM breaks down your input into tokens. Tokens are individual units of text that represent words, phrases, or other meaningful elements. The number of tokens in your prompt or response might vary depending on the complexity of the task and the length of the input.
The amount of information (context length) that AI agent considers when it processes a user's prompt varies depending on the base LLM that the AI agent uses.
By default, it uses the IBM Granite model, but tenant administrators can change this default by configuring the model of the AI agent.
Mentioning AI assistants, agents, and skills
Navigate the conversation straight to the AI assistants, agents, or skills that are specialized in the domain required to complete a task. Mention the AI assistants, agents, or skills' name with the @ symbol and include your prompt
in the chat bar. For example, "@RetailAnalysis show me the sales performance by product category for the past month."
Use the mention to ensure the most appropriate AI assistants, agents, or skills are dealing with a specific issue, potentially improving efficiency and accuracy.
Routing based on the conversation context
Your conversation is automatically directed to an AI assistant, AI agent, or skill based on the conversation context. The conversation history forms the context of the chat, your current input, and the specific details of the AI assistants, AI agents, and skills added to the chat.
To ensure seamless execution of AI assistants' actions, the agent routes with an AI assistant for all subsequent interactions until the assistant completes its actions. It does not suggest new routing recommendations during an ongoing conversation. The AI assistant ends automatically in the following scenarios:
- All Action Completion: The AI assistant ends after it completes all its designed actions. For more details, see Finishing the AI assistant.
- Direct invocation: The AI assistant ends when directly started by using mentions.
- Conversational search response: The AI assistant ends after Conversational search content responds.
Finishing the action
To finish an action, refer to Subactions. You can create multiple actions and design one action to start another using subactions based on your requirements.
Finishing the AI assistant
The AI agent allows multiple AI assistants to work together. The underlying LLM router routes messages to the best-matching assistant and continues until the assistant completes its tasks. To ensure the assistant ends its operations smoothly, follow these best practices:
| Practice | Action |
|---|---|
| Action completed | Use End the Action in the And then section for any required step. |
| Subactions completion | To start a subaction, use the End this action after the other action is completed checkbox, or use End the Action in a parent action step. |
| Custom fallback | Review custom fallback connections to Agents or external flows. Decide whether the agent can reroute more effectively to other assistants. |
| Disambiguation and digression | Revisit configurations for Ask clarifying questions or Change conversation topic features. If they are unnecessary or better handled by another assistant, consider disabling them. |
Viewing your tasks
A task is a skill or skill flow that you run, or an LLM task that is being generated. If the task you requested takes time to run, the AI chat suggests you work on something else while the requested task runs in your chat. You can start more than one skill over multiple chats. You receive a notification with the task status and task completion.
You can view your recent tasks or chat threads in the left pane of the AI chat. The chat threads under the Recent chats section are sorted by time, which helps you to track the chat with latest updates. Even if you have an unread chat, the system displays the chat thread with the latest update on the top.