The development tools available to achieve those four types of characteristics (pleasing persona, clear objectives, algorithms, and constraints) are context management, prompt articulation, bridge building, and Large Language Model (LLM) selection.

Context Window

GenAI LLM’s typically have two ways to create input: prompt creation and loading of the context window. The context window is ‘background data’ that the LLM uses in conjunction with specific prompts to produce inferences. Fully specifying a personality can be done in a prompt. However, it may be easier to get the desired results when creating a personality when there is a very detailed description in the context window. 

Giving that personality description a name that can be routinely referenced can also be helpful. If this personality approach is used, there may be an option of informing the customer that they are interacting with AI. Some environments may require such notification. In environments that don’t require it, there is a choice to be made.

Many LLM’s deliver the results of each inference session into their context window. This can amplify errors and increase the probability of hallucinations or inaccurate results. So, consideration should be given to how to manage the context window over time. There are many strategies ranging from erasing and reloading the base (including personality) context window at the end of each inference session, to not deleting any information in the context window until it hits an overflow threshold and then allowing the algorithm the LLM uses to make space.

Prompts

The prompt describes to the LLM what is desired. The more explicit and detailed the prompt is, the better the result. The prompt doesn’t have to be written in a programing language, Boolean logic statement, mathematical equation, or other structure, though those types can be used. The fundamental objective is to make it as clear as possible to the LLM what the desired output should be.

The best way to do this for an intelligent agent is to describe the Objectives, Algorithms, and Constraints for the agent.

Objectives for a help desk can be generic, and some will be specific. For example, a generic objective might be something like, “Always try to make the customer comfortable.” Or “Always try to understand exactly what the customer is saying.” Domain specific objectives generally have to do with key aspects if the domain.

Algorithms are ways of describing processes that the intelligent agent can or should use to achieve the objectives. Some examples of algorithms include the following. Using the data entry scripts described above. Repeating back to the customer in slightly different wording than what the customer says and asking if that accurately reflects what the customer was trying to address. Some algorithms may be mathematical or Boolean (if then, else chains, etc.). In the Boolean case, one set of algorithms may include conditions where the agent transfers the session to a human staff member.

Constraints are limits on the action of the intelligent agent. Examples might include the following. “Never express, or appear to express, frustration with the customer.” “Never quote a price over $500”. “Never say that something is the organizations’ fault.” etc. Constraints may also be expressed in Boolean terms.  Constraints may be expressed mathematically in some pricing or terms of conditions discussions or in highly technical domains.

Bridges and Umbrella Data Model

Bridges are the subsystems that intelligent agents use to connect with other subsystems that they obtain information from, send information to, or seek to control. The bridge provides data translation, API implementation, etc. to do this job. There may be many bridges because organizations often have sub systems that came from different generations of technology, cultural contexts, and ways of representing data.

It is generally helpful to have an overall data model with standard terminology. This overall data model is called an Umbrella Model. The Umbrella Data Model is a superset of all of the local data models. The bridge translates from/to the Umbrella Model and the local data models. An example of such is the use of Fahrenheit and Celsius measures of temperature. The Umbrella Model may represent all temperatures in Fahrenheit while some sub systems local data model represents temperature in Celsius. In such a situation, the bridge would translate Celsius into Fahrenheit for the agent and then back to Celsius for the sub system.