Skip to main content

Advanced API

In addition to the fundamental methods new CortexStep(name), withMemory, and next, the CortexStep class offers a set of advanced capabilities to gain more control over the behavior of the AI's internal monologue. This page explains how to use these advanced features.

Full Constructor Signature​

The CortexStep class accepts an optional CortexStepOptions object in the constructor. This allows you to provide past steps, a processor, initial memories, or a last generated value.

constructor(entityName: string, options?: CortexStepOptions)

The CortexStepOptions object includes:

  • pastCortexStep: A CortexStep instance representing the previous step.
  • processor: A LanguageModelProgramExecutor instance handling the execution of language model instructions.
  • memories: An array of WorkingMemory instances representing initial memories.
  • lastValue: The last generated value.
  • id: a unique identifier of this particular cortext step
  • parents: used to keep track of previous cortext steps in a chain
  • tags: keeps tags on a cortex step that are added to instrumentation calls.
let step = new CortexStep("Assistant", {
  pastCortexStep: previousStep,
  processor: new OpenAILanguageProgramProcessor(),
  memories: [[{ role: "system", content: "You are a helpful assistant." }]],
  lastValue: "Solved a complex math problem",
});

Other Methods​

toString()​

toString() generates a string representation of the assistant's chat history, including system instructions, user queries, and assistant responses.

let stringRepresentation = step.toString();

Understanding and Utilizing The Value Property​

CortexStep has a value property which represents the output generated from the last action. This value is strongly typed depending on the evaluated Cognitive Function.

This value can be very useful in chaining actions where the output of one action might influence or be used in the next action. To access this value, simply use:

let value = step.value;

Remember, the CortexStep class is designed to be functional - that is, none of its methods alter the state of the existing instance. Instead, they return a new instance with the updated state. This design allows for easy chaining of actions and states while maintaining immutability.