Introduction to CALM
Video
The rise of Large Language Models (LLMs) has enabled developers to create highly interactive, open-ended assistants capable of handling a vast range of requests and elevating the conversational quality of these interactions. However, assistants built solely using LLMs are prone to hallucinations, which are extremely challenging to control in a production setting. While such hallucinations can be entertaining in casual environments, they can be costly and problematic in enterprise-grade use cases like customer support.
Enterprise-level assistants are built with specific purposes in mind, such as customer support, airline ticket booking, or other targeted tasks. Their true value lies in their ability to efficiently meet users' needs, often involving business logic—specific rules and regulations that must be followed. In an enterprise setting, it is crucial that assistants adhere strictly to this business logic.
The power of CALM
Rasa’s CALM (Conversational AI with Language Models) technology addresses this by combining the best aspects of LLM dialogue management with easily programmable logic. This approach merges the fluency and generality of LLMs with the precision of NLU-based systems, empowering developers to create conversational AI assistants that are effective, engaging, and straightforward to build. With CALM, there’s no need to create extensive training examples for conversations; instead, the focus is on defining business logic and building the assistant.
Key components on CALM
CALM operates through three core components: business logic, dialogue understanding, and conversation repair.
Business Logic: This component serves as the backbone, outlining the necessary steps to fulfill a user's request. CALM implements business logic as a set of flows that describe the required information from the user, any data retrieval from APIs or databases, and branching logic based on the information collected. Essentially, it comprises rules that guide the assistant in specific situations to successfully complete predefined tasks. For instance, in a banking customer support assistant, a business logic rule might dictate that if a user reports a lost card, it should be blocked immediately.
Dialogue Understanding: This system ensures that the assistant comprehends user inputs while maintaining the context of what the user aims to achieve. It involves generating commands that reflect the user's intentions, aligned with business logic and the conversation's context. Commands exist for starting and stopping flows, filling slots, and more, ensuring that users can engage in natural conversations while progressing toward their goals.
Conversation Repair: When users pose questions or scenarios outside the assistant's scope, conversation repair steps in. Its role is to gracefully redirect users back to the goal-oriented conversation. Whether handling chit-chat, seeking clarifications, or addressing queries outside the designated scope, conversation repair enables assistants to navigate unexpected situations smoothly.
By leveraging CALM, developers can create conversational AI assistants that are not only powerful and engaging but also adhere strictly to business logic, ensuring reliability and efficiency in enterprise applications.