Implementing
How to Start
Once you have a good first-cut idea of the general conversational design of your agent, it’s a good idea to start exploring it by actually implementing it, piece by piece.
You will find that as you discover the strengths and weaknesses of your approach, you will need to circle back and modify your design based on real world interactions.
iostack allows you to do this easily: adding, testing and modifying all aspects of your agent can be done entirely within the Agent Flow and the other Settings sections.
Iterating
A conversational process is often conceived of as a ‘golden path’ i.e. an idealised process through your conversational structure.
You should focus on implementing this initially - but also keep in mind that conversations will often not follow this path in the real world.
The ability to deal with these alternative flows will determine the degree to which your agent is truly flexible, coherent and nuanced in its delivery of your process.
Options
iostack’s ability to exploit the intelligence of the LLM in a constrained and structured way gives you a few alternatives:
-
Using prompting, you could leverage the intelligence and flexibility of the LLM to coerce your users through the golden path even if they may digress or want to talk about trivia ie. force them through your predefined process, utilising the intelligence of the LLM to maintain coherence and to counter objections.
-
Alternatively, you can easily add additional transitions at key points in the conversation structure to allow the user to change their mind. You could also choose to expose key variables for update by the LLM across a wider range of stages, allowing values to be updated in a nuanced way as the conversation and the process evolve naturally and it becomes clearer what the user is wanting. These approaches may have to be done in conjunction with prompt changes to allow the conversation to maintain coherence in the case of revisiting an earlier conversational stage.
-
Or you could experiment with making the relevant stages of a more fluid conversational process maximally interconnected i.e. moving from any stage to any other other is valid, allowing the user to jump discursively between related but not strictly ordered stages while evolving a set of universally interrelated variables exposed at all stages.
This is by no means an exhaustive list of possibilities - our experience has been that the combination of a constraining structure and exploiting the intelligence of the LLM rather than relying on rules or heuristics has opened up a potentially unlimited number of approaches to modelling various conversational structures.
We’re super interested to see what can be realised using iostack – so be sure to let us know what you’ve built!