5 Conversational AI Frameworks for AI Agents

Key Takeaways

last episode I touched on the future of conversation design and in that I introduced the idea of conversation Frameworks I had a few people message me after seeing the episode and say hey these Frameworks sound cool but Pete what the [ __ ] are they so let's chat about that hey it's Pete and today we're delving into the future of conversation designs and how Frameworks will be the backbone of what we create so first up what is a framework well if you think of your agent as a new employee then the framework is like their handbook it instructs them on how to behave respond and navigate complex interactions and here's what I think they'll look like one we've got Global Frameworks the overarching rules that an agent will work within two local Frameworks the tailored strategies for managing specific conversation States three interaction Frameworks the essence of communication style tone and technique four data handling and analytics Frameworks how we pass information to the language model and how we derive insights from our interactions and five integration Frameworks the bridges that we build for our agent to access external systems now I don't have time to focus on all five so in this episode we're going to focus on the first two which I suspect conversation designers or agent designers that I alluded to in the last episode will focus on the most so let's start with global Frameworks they include things like routing logic error recovery context management and to help with a mental model of how they might work all together let's use an example of a person planning a trip to Tokyo thanks to the routing framework the agent knows the user is asking about booking a trip and places them in that conversation State the user is thinking about going in August but instead of saying that explicitly they ask what's the weather like in August the routing logic enables the agent to pull this information while the error framework works on getting the user back on track asking if August works or maybe they'd like something cooler this is where the global framework's context management kicks into gear pulling information from prior trips offering their preferred seat type which is window along with their preferred Airline VF I mean whose's wouldn't be okay so if Global Frameworks are the rules that an agent Works within then local Frameworks are the specific states that an agent can access you can kind of think of them as like business processes let's look at a situation that we've actually been helping customers with a lot designing Frameworks for retrieval augmented generation that minimize errors and keep true to many of the cxd best practices that we use today here's what the customer's response looked like without a framework built around it on the surface this doesn't seem bad but when you're reading it you start to realize this response while written well doesn't really tell you much it doesn't mention a device the instructions are kind of vague and it doesn't actually give you any APN information to enter to be fair it's kind of useless so here's what we do upon entering the framework we look at the conversation history thus far and write a question that's optimized for retrieval augmented generation as you can see that llm has added to my device to make the question a little more pointed this should help with the retrieval side of things but the question is still pretty vague next we instruct the llm to look at the information it has retrieved along with the question that it has presented to the knowledge base and ask could we get a more pointed answer if we got more information from the user here you can see that the llm has determined that the user needs to specify the device in order to give them the most pointed information the user responds with iPhone 15 and is asked for further information so that it can generate the final question for the knowledge base which is the following now we have all the information the response is created but before it's presented to the user the llm checks there are no inconsistencies or hallucinations by doing a cross check of the response and the information that was retrieved finally the answer is displayed and another prompt is used to display a follow-up question which it asks to the user this local framework gives llms the tools to create questions that are optimized for retrieval ask clarifying questions selfcheck its own work and ask followup questions to keep a conversation going and using it you can go from managing hundreds of FAQs to managing a knowledge source that the framework accesses and this is really what conversation Frameworks are all about they are or will be I believe the backbone of every llm based agent they will set the overarching rules that an agent could work within the different conversational States it can access the communication style and tone it will use how it will handle specific data and what types of Integrations it will use and when it will use them depending on the current state of the conversation okay for everyone that has stuck through me to the end thank you and that my friends is my take on the future of conversation design and how we'll use Frameworks to Glide agents through complex conversations let me know your thoughts in the comments and remember stay curious [Music]