Skip to main content

Core API

The CortexStep class is a critical component of the SocialAGI library, designed to provide a structured way to model the cognitive process of an AI during a conversational interaction. This is particularly focused on simulating an AI's internal monologue or thought process, capturing how an AI interprets input, makes decisions, and crafts responses.

The class adheres to the principles of functional programming. Each method in CortexStep generates a new instance, preserving immutability and enabling a more predictable behavior.

Key Concepts and Methods​

  1. CortexStep Initialization

    To initialize CortexStep, you provide an entityName. This refers to the AI's identity.

    let step = new CortexStep("Assistant");

  2. Building Memory with withMemory()

    Memories are built using ChatMessage objects, each representing a discrete cognitive event or "step". withMemory() adds a memory to the AI's existing set.

    step = step.withMemory([{ role: "user", content: "Hello, Assistant!" }]);

  3. Next Cognitive Step with next()

    The next() method guides the AI to the subsequent cognitive step. This method requires a cognitive function. The library offers built in cognitive functions, or you can build your own.

    step = await step.next(internalMonologue("thinks, understanding the user's message"));

  4. Get only values with compute()

    The compute() method works just like next() but does not provide a new CortexStep and instead just returns the value from the CognitiveFunction

    const yesOrNo = await step.compute(decision("Is Bogus cold?", ["yes", "no"]));

  5. Resetting Memory with withUpdatedMemory()

    The withUpdatedMemory() method allows the AI to reset its memories based on a function that processes the existing memories. It returns a new CortexStep instance with the updated memories.

    step = await step.withUpdatedMemory(memories => memories.filter(m => m.role !== "user"));

Predefined CognitiveFunctions​

While CortexStep allows for the creation of custom cognitive functions, it comes with several predefined cognitive functions:

  1. internalMonologue: Simulates the AI's internal thought process.

    step = await step.next(internalMonologue("Contemplate the users message"));

  2. externalDialog: Creates an external dialogue or output from the AI.

    step = await step.next(externalDialog("Give the user a greeting message"));

  3. decision: Guides the AI through a decision-making process.

    step = await step.next(decision(`Choose the next topic`, ["Weather", "News", "Sports"]));

  4. brainstorm: Facilitates brainstorming of potential actions for the AI.

    step = await step.next(brainstorm("Brainstorm the next conversation topics"));

Generated output​

value retrieves the value generated from the last action.

let lastValue = step.value;

Streaming support​

Response streaming is fully support on .next steps

const { stream, nextStep } = await step.next(internalMonologue("thinks, understanding the user's message"), { stream: true });
for await (const chunk of stream) {
   console.log(chunk);
}
step = await nextStep;

Model specification​

By default, CortexStep is configured to use OpenAI models. On each .next step or .compute call, an OpenAI model string can optionally be passed as .next(..., { model: "gpt-3.5-turbo-1106" }).

We also have first-class support for open source models through optional specification of a language model executor interface when constructing a CortexStep.