The NAF contains an observer system that functions as a parallel reasoning system running alongside the main conversational agent. Think of it as a "coach" that watches the conversation and continuously provides guidance (thoughts) to keep the agent on track. This thought process guides conversation flow to prevent missed steps and hallucinations.
Functionality
The Problem: Without guidance, LLMs can "wander" and skip steps or jump to wrong conclusions.
The Solution: The Observer tells the agent exactly where it is and what comes next.
Observes the conversational history and original prompt
Analyzes the entire conversation context including:
Conversation metadata (e.g., business hours)
Business context for accurate responses
All scenarios and procedures
Generates structured internal outputs (thoughts) after each agent turn
Example Thoughts Output
Example Workflow
User: "I would like to reserve a table for 15 people for this Sunday"
Observer analyzes and identifies: Large group reservation scenario
Observer generates key directives: "Proceed to step 3.2 - Find out user's phone number and email address"
Agent follows directed path instead of potentially missing steps, thereby, drastically reducing the chance of a hallucination.