Powered by

# Concepts

# System Prompt

The System Prompt is a persistent set of instructions that defines the role, purpose, tone, behavior, and constraints of a Digital Expert.
It serves as the foundational guidance that the model follows for the entire interaction, influencing how it interprets user requests and delivers responses.

In Purple Fabric, the System Prompt is configured through the Perspective field and typically includes:

  • Role: The specific identity of the Digital Expert or Team of Agents.
  • Persona: The communication style (e.g., formal, concise, customer‑friendly).
  • Goals & Responsibilities: The tasks the Digital Expert is expected to perform.
  • Guidelines: Formatting, tone, and content rules.
  • Outcome Presentation: Preferred format for presenting results (e.g., bullet points, tables, summaries).
    This prompt is not visible to the end‑user but affects every response.
    Example: “You are a financial compliance Digital Expert for a retail bank. Explain regulatory requirements in clear, concise language, avoid unnecessary jargon, and where possible, provide relevant banking examples. Present findings in bullet points with key takeaways.”

# User Prompt

The User Prompt is the immediate request, question, or instruction given by the user during an interaction.
It changes with each query and is interpreted in the context of the System Prompt.
While the System Prompt defines how the Digital Expert should respond, the User Prompt specifies what the Digital Expert should respond to at that moment.
Example: “Summarize the key changes in Basel III banking regulations for 2025 and explain how they affect Tier 1 capital requirements.”

Tokens
The fundamental unit of text processing for Large Language Models (LLMs) within Purple Fabric. A token can be a word, part of a word, or punctuation, used to measure the length of prompts and responses. Tokens are the basis for managing LLM performance and cost. Tracking usage ensures agents operate efficiently within predefined limits, preventing unexpected costs.
Pre‑requisite: Tokens are an inherent part of LLM processing. They are automatically consumed during every agent interaction.
Example: A simple query might use 5 tokens, while a detailed agent response could use 100-200.

# Models

Models are the foundational AI components that enable your agents to comprehend, reason, and generate responses. You select and configure these models for specific task requirements.

  • Large Language Model (LLM)
    A LLM is the “brain” of your AI agents, a powerful system that understands and generates human language. It enables agents to interpret context, infer meaning, and respond naturally to user inputs.
    Example: In Purple Fabric, an LLM powers a customer support agent to read a user’s query, understand the intent, and craft a relevant, conversational reply in real time.
  • Embedding Models Embedding models turn text into numeric patterns that capture meaning.
    To better understand refer the Embedding page
  • Re-ranker A reranker is a component in an advanced search system that refines the initial list of results. To better understand, refer to the Re-ranker page.

# Prompt Engineering

Prompt Engineering is the strategic process of crafting precise inputs for your Digital Experts.

  • Purpose: Guides the agent towards desired outputs and behaviors.
  • Benefit: Optimizes agent understanding and response quality, preventing undesirable outcomes.
  • Perspective
    • Definition: Foundational directives defining the core identity, persona, and operational guidelines.
    • Configuration: Define the role, tone, and ethical boundaries.
    • Impact: Shapes prompt interpretation and response style for consistency.
  • Experience
    • Definition: Curated input-output pairs demonstrating desired business outcome.
    • Purpose: Fine-tunes the understanding of complex queries.
    • Benefit: Improves consistent and accurate performance in real-world interactions.

# AI Agent

AI Agents are autonomous AI programs performing tasks based on their configurations.

  • Configuration: You configure them with models, knowledge sources, and tools.
  • Capability: Understand prompts, process information, and complete specific tasks.

# Tools

Tools in Purple Fabric include external applications, connectors, and utilities that AI agents can use to extend their capabilities.
For detailed definition, refer to the Tool page.

# Digital Expert

A Digital Expert is an intelligent AI‑driven solution designed to perform specific business functions and handle compound tasks that require multi‑step reasoning or actions.

  • Purpose: Provides specialized knowledge or automates complex operational processes.
  • Role: Acts as a dedicated, virtual partner, augmenting human capabilities.
  • Focus: Tailored to specific domains for deep expertise.

# Team of Agents

A Team of Agents is a collaborative group of individual AI agents.

  • Purpose: Tackles complex, multi-faceted business objectives.
  • Collaboration: Each agent contributes specialized capabilities.
  • Mechanism: Works through defined communication protocols and handoff mechanisms to achieve shared goals.

# Retrieval-Augmented Generation (RAG)

RAG enhances an LLM’s knowledge with reliable, verified, and domain‑specific data from trusted sources.

  • Mechanism: "Grounds" agent responses in your verified enterprise knowledge base.
  • Benefit: Significantly reduces "hallucinations" and improves factual accuracy.
  • Outcome: Provides relevant, context‑rich information to improve the accuracy and trustworthiness of outputs.

# ReAct Framework

The ReAct Framework combines Reasoning with Acting capabilities in a team of agents working together.

  • Function: Enables the team to plan and adjust sequences of decisions and actions dynamically.
  • Capability: Collaborates intelligently, uses tools, and retrieves targeted information when needed.
  • Result: Forms a decision‑driven, flexible, and adaptive problem‑solving loop.
    Example: In a loan approval workflow, one agent verifies customer documents, another checks credit history, and a third applies lending rules. The team decides the next steps based on each result, ensuring accurate and compliant approvals.

# Prime Agent Framework Overview

The Prime Agent framework represents a significant advancement in conversational AI development, providing a centralized and efficient architecture for building versatile conversational agents.

1. Unified Conversational Agent Framework
The Prime Agent is a framework designed to provide a unified window for creating, configuring, and deploying conversational agents. It consolidates the necessary features into a single environment, streamlining the development process.

2. Lifecycle Management and Deletion
For effective agent management, the Prime Agent offers straightforward lifecycle control. Agents built using this framework can be easily selected directly from the Agent Template dropdown menu.

3. Deployment Flexibility: Single Agent and Team of Agent
The Prime Agent framework supports highly flexible deployment models, allowing it to be used in two primary configurations:

  • Single Agent: The Prime Agent operates independently to handle user queries.
  • Team of Agents: The Prime Agent acts as the orchestrator, managing the flow of conversation across multiple specialized agents.

4. Single Agent Capabilities In the Single Agent configuration, the Prime Agent maintains robust capabilities, supporting integrated knowledge resources for enhanced performance:

  • Tool Integration: The ability to execute external functions or operations.
  • Knowledge Garden (KG) Integration: Direct access to structured knowledge bases for grounded responses.

5. Team Agent Orchestration
When configured as a Team of Agents, the Prime Agent expands its capabilities to manage complex, multi-faceted interactions:

  • Tool Integration
  • Knowledge Garden (KG) Integration
  • Sub-Agent Inclusion: The ability to orchestrate and delegate tasks to specialized sub-agents.

6. Performance Distinction: Rule-Based Execution vs. ReAct
The Prime Agent architecture is fundamentally different from frameworks utilizing the ReAct (Reason and Act) approach.

  • ReAct: Focuses on dynamic reasoning and action planning at each turn, which can introduce latency.
  • Prime Agent: Follows pre-defined rules in turn, allowing for significantly faster execution times relative to the dynamic planning inherent in ReAct models.

# Different Templates for Building AI Agents

This section outlines the primary templates available for configuring and deploying AI Agents, categorized by their structural type and core functionality.

1. Default Agent

  • Type: Atomic.
  • Function: A simple, foundational AI Agent designed for executing atomic tasks.
  • Mechanism: Direct execution of the user prompt, relying solely on the Large Language Model's (LLM) base training knowledge.
  • Benefit: Provides the lowest execution overhead, resulting in the fastest response time for simple, non-grounded queries.
  • Outcome: A direct, single-step response with no external information retrieval or tool interaction.
  • Example: Answering a commonly known factual question, such as "What is the capital of Japan?"

2. Retrieval-Augmented Generation (RAG) Agent

  • Type: Atomic.
  • Function: Enhances the LLM’s knowledge by incorporating reliable, verified, and domain-specific data from trusted enterprise sources.
  • Mechanism: Grounds agent responses in your verified enterprise knowledge garden through targeted retrieval before generation.
  • Benefit: Significantly reduces the occurrence of model "hallucinations" and improves the factual accuracy and trustworthiness of outputs.
  • Outcome: Provides relevant, context-rich information, ensuring generated content is directly supported by verified data.
  • Example: Summarizing the terms and conditions from an internal company legal document based on a user's specific query.

3. ReAct Agent

  • Type: Molecular.
  • Function: Combines Reasoning with Acting capabilities within a collaborative team of agents.
  • Mechanism: Enables the team to plan and dynamically adjust sequences of decisions and actions based on intermediate results, often involving intelligent collaboration and targeted tool utilization.
  • Benefit: Highly adaptive, capable of handling complex, multi-step problems that require dynamic decision-making and tool execution.
  • Result: Forms a decision-driven, flexible, and adaptive problem-solving loop through intelligent collaboration, tool utilization, and targeted information retrieval.
  • Example: In a loan approval workflow, specialized agents collaborate to verify documents, check credit history, and apply lending rules, collectively deciding the next compliant step based on real-time results.

4. Prime Agent

  • Type: Atomic or Molecular & Description: Represents a significant advancement in conversational AI, providing a centralized and efficient architecture for building versatile conversational agents.
    4.1. Architecture and Deployment

    • Deployment Flexibility: Supports highly flexible deployment models, including Single Agent (independent operation) and Team of Agents (acting as the orchestrator).
    • Unified Conversational Agent Framework: Provides a single window for creating, configuring, and deploying agents, streamlining lifecycle management.

    4.2. Core Capabilities

    • Mechanism: Executes tasks based on a framework focusing on efficient flow control and task delegation when acting as an orchestrator.
    • Integrated capabilities across all modes: Tool Integration and Knowledge Garden (KG) Integration.
    • Benefit: Allows for significantly faster execution times relative to the dynamic, latency-introducing planning inherent in ReAct models.
    • Outcome: Provides a versatile agent capable of handling complex interactions, maintaining high performance, and utilizing integrated knowledge resources and sub-agents.
    • Example: Orchestrating a full customer support interaction, where the Prime Agent delegates an account lookup task to a Sub-Agent where it can also be a Sub- Agent and then uses its own EKG integration to provide a solution.

# Agent Configurations

# Models

You configure the models that provide computational intelligence for your Digital Expert.

  • Selecting LLM Models
    • Function: Select foundational LLM.
    • Impact: Determines the agent's general knowledge, reasoning, and language generation.

# Lookback Limit


Defines how many previous conversations can be referenced from the history when generating a response to ensure the contextual continuity.

  • Default: 10
  • Range: 0–50
  • Example: If set to 5, the agent will only consider the last five interactions for context.

# Model Settings

# Temperature

Controls the creativity vs. predictability of the model’s responses.

  • Lower values: More focused, predictable, and consistent outputs.
  • Higher values: More creative, varied, and less deterministic responses.
  • Range: Typically 0–2 for GPT models; 0–1 for certain other models.
  • Example: 0.2 is ideal for compliance documentation; 1.0 is better for brainstorming ideas.

# Top‑P:

Controls how widely the AI considers different wording options when creating a response.

  • Lower values: The AI chooses from only the most likely words, giving more predictable and consistent answers.
  • Higher values: The AI considers a wider range of possible words, which can make responses more creative but less predictable.
  • Range: 0–1
  • Example: 0.9 allows diverse responses; 0.3 narrows to safer, more predictable answers.

# Logprobs

Displays the model’s confidence score for each generated word, showing how certain the model is about its output.

  • Availability: Only supported on Azure GPT‑4o models:
    • Azure GPT‑4o‑mini
    • Azure GPT‑4o‑4K
    • Azure GPT‑4o‑16K
  • Use Case: Debugging agent responses, fine‑tuning prompts, and analyzing model decision‑making.
  • Control: On/Off toggle.
  • Example: When enabled for a supported Azure GPT‑4o model, the system displays probabilities for each generated word, helping identify uncertainty or hesitation in the model’s responses.

# Advanced Retrieval Settings

# Top‑K

Specifies how many pieces of information (“chunks”) the system initially retrieves from the Enterprise Knowledge Garden (EKG) or other connected knowledge sources when answering a query.

  • If re‑ranking is off, all of these go straight to the model.
  • Vision models can still only work with up to 20 images, even if Top‑K is set higher.
  • Range: 1-100
  • Example: If you set Top‑K to 10, the system will pull the 10 most relevant chunks it can find.
# Re‑ranker

Helps the system sort the Top‑K results so the best ones are used.

  • If enabled, only the most relevant chunks are kept for the final answer.
  • Toggle: On/Off
  • Example Model: Azure AI Cohere Rerank v3.5
# Top‑N

How many chunks from the Top‑K should be re‑ranked.

  • Range: 1–50
  • Example: If Top‑K is 10 and Top‑N is 5, the system will keep the 5 best chunks from the original 10.
# Self‑Query (Toggle)

Enables the system to generate smarter, metadata‑driven queries to improve retrieval accuracy and surface more relevant results.

  • Pre‑requisite: The Enterprise Knowledge Garden (EKG) must contain chunks with properly tagged metadata (e.g., author, date, category) for Self‑Query to work effectively. Without metadata tags, the feature cannot optimize retrieval.
  • Benefit: Helps the agent filter and target information more precisely, improving the relevance of responses.
  • Example: If EKG chunks are tagged with “region” and “year”, the system can automatically refine a user query like “Show sales trends” into “Show sales trends in APAC for 2024” using metadata filters.
# Filtered References

This feature filters your sources to highlight only those that directly contributed to the generated response. It is optimized for high-performance models such as GPT-4o, GPT-4o-mini, and Claude 3.5 Sonnet.

# Research Mode (Toggle)

Slows things down slightly to dig deeper and be more precise — useful for complex questions or detailed analysis.

# Research Mode Prompt

Special instructions you can set to guide the model when Research Mode is on, ensuring it focuses on exactly what you need.

# Interaction Types

You define how your Digital Experts will engage with users or systems: conversational dialogue or automated execution.


Conversation
Enables Digital Experts to engage in natural language dialogues with users, supporting multi‑turn interactions where context is maintained across multiple exchanges.

  • Function: Facilitate fluid, human‑like conversations that adapt to the user’s inputs over time.
  • Application: Customer support agents, virtual assistants.

Automation Enables Digital Experts to use input/output (I/O) schemas to define the structure of data they receive and produce, ensuring reliable integration with other systems and processes.

  • Function: Perform actions autonomously by processing structured inputs and delivering outputs in consistent, expected formats.
  • Application: Invoice Extractor, KYC Classifier, Claims Verification Expert.

Add-ons are optional, configurable modules that extend your Digital Expert’s features beyond its primary functions.

# Automation Agent Variable Definition

1. Defining Dynamic Input Variables
When creating an Automation Agent, dynamic values must be registered at the agent level:

  • Variable Definition: Input variables intended for dynamic execution must be explicitly defined within the parameters section of the automation agent's configuration.

2. Referencing Input Variables (Single Braces)
To successfully integrate a defined input variable into the agent (e.g., within the logic of a Perspective), a specific referencing syntax must be used:

  • Referencing Syntax: To utilize a previously defined variable, the variable name must be enclosed within single curly braces ().
    • Example: If the input variable is named customerID, it must be referenced as .
  • System Recognition: The system only recognizes the variable as a dynamic reference when it is correctly encapsulated inside these single curly braces.

3. Formatting Structured JSON Output (Double Braces)
When the expected outcome of the automation sequence requires a structured data format, specifically JSON, a distinct encapsulation method is required to ensure proper parsing:

  • JSON Output Requirement: If the desired output structure is JSON (JavaScript Object Notation), the entire JSON payload must be encapsulated within double curly brackets (). This ensures the system interprets the content as a structured object for subsequent processing.