Compliance Advisor Bot

​

1. Agent Creation and Core Definition

​

Step 1.1: Accessing the Studio and Initiating Agent Creation

1. Log In: Navigate to the Lyzr Studio platform and log in.
2. Navigate to Agents: In the main dashboard, select the Agents tab.
3. Create New Agent: Click the + Create Agent button, located in the top right corner.

​

Step 1.2: Defining the Agent’s Core Identity

This involves setting the agent’s name, description, and selecting the foundational LLM.

1. Name and Description: Fill in the agent’s Name (e.g., “Compliance Advisor Bot”) and Description (e.g., “You are a highly professional and diligent Compliance Advisor”).
2. Model Selection: Choose the Provider (e.g., OpenAI, Google) and select the specific Model (e.g., gpt-4o-mini).

3. Agent Role and Goal: Define the Agent Role (e.g., “You are an expert customer support agent”) and the Agent Goal (e.g., “Your goal is to address and resolve customer inquiries”).

​

Step 1.3: Setting the Direction (Agent Instructions)

Provide detailed, strict instructions in the Agent Instructions field to enforce the role and anti-hallucination guardrails.

• Example Focus: The instructions direct the agent to LISTEN to the customer’s concern & GATHER all relevant information needed for resolution. PROVIDE clear and concise answers to the customer.

​

Step 1.4: Configuring Contextual Memory

This ensures the agent can maintain the flow of conversation.

1. Enable Memory: Toggle the Memory feature switch under Core Features.
2. Configure: Click the gear icon or configuration link next to the Memory feature.
3. Set Parameters:
   • Choose the Memory Provider (e.g., Lyzr Memory).
   • Set the Max. messages stored as Short-term Memory using the slider (e.g., set to 10 messages for robust context).
4. Save: Click Save to apply the configuration.

​

2. Knowledge Base (RAG) Setup

This process grounds the agent’s answers in your organizational data, addressing the “Groundedness” Responsible AI concern.

​

Step 2.1: Initiating Knowledge Base Creation

1. Navigate to Knowledge Bases: Go to the dedicated Knowledge Base section.
2. Create New: Click the + Create New button.

3. Select Type: Select the Knowledge Base option, which is suitable for quick answers from documents.

​

Step 2.2: Defining the Knowledge Base Structure

1. Name and Description: Provide a unique Name (e.g., hr_policies) and a Description (e.g., Compliance Data).
2. Select Components:
   • Vector Store: Choose the database provider (e.g., Weaviate [Lyzr]).
   • LLM Embedding Model: Select the model used for creating vectors (e.g., text-embedding-3-large).
3. Save Structure: Click Create.

​

Step 2.3: Uploading and Processing Content

Upload the actual source files that the agent will use for factual answers.

1. Upload Files: Drag and drop or click to select your official documents (PDF, DOCX, TXT files).
   • Example: Upload your policy document (e.g., CHEMEXCIL_HR_Policy_2024_08-01-2024.pdf).
2. PDF Parser: Select the appropriate parser (e.g., LLMSherpa).
3. Continue: Click Continue to start the ingestion and processing.

​

Step 2.4: Attaching the Knowledge Base to the Agent

1. Return to Agent Builder: Go back to the configuration page for your “Compliance Advisor Bot.”
2. Enable Knowledge Base: Toggle the Knowledge Base switch under Core Features.
3. Configure RAG: Click the gear icon.
4. Select Knowledge Base: Select the created knowledge base (e.g., hr_policies).
5. Set RAG Parameters:
   • Number of Chunks: Set the number of retrieved document sections passed to the LLM (e.g., 7).
   • Retrieval Type: Select the retrieval algorithm (e.g., MMR).
   • Score Threshold: Set the minimum relevance score (e.g., 0.5) for chunks to be used.
6. Save: Click Save.

​

3. Creating the Responsible AI Policy

The Lyzr Studio uses centralized policies to manage guardrails across multiple agents. We must create this policy first.

​

Step 3.1: Navigate to Responsible AI and Initiate Policy Creation

1. Navigate: From the main Lyzr dashboard, click on the Responsible AI tab (or equivalent section).
2. Create New: Click the + Create New button in the top right corner.

3. Define Policy: A “Create new policy” modal appears.
   • Policy Name: Enter a name like HR_Compliance_Policy.
   • Description: Add a brief purpose, e.g., Enforces PII Redaction and Content Moderation for internal HR agents.
   • Create: Click Create.

​

Step 3.2: Configuring Core Safety Guardrails (Toxicity & Prompt Injection)

These settings protect the agent from misuse and maintain a professional environment within the policy configuration screen.

1. Toxicity Controller:
   • Activate: Toggle the Toxicity switch ON.
   • Threshold: Adjust the slider (e.g., set to 0.3 or lower) to monitor and prevent toxic or harmful content.
2. Prompt Injection Manager:
   • Activate: Toggle the Prompt Injection switch ON.
   • Threshold: Set a strict Threshold (e.g., set to 0.3 or lower) to protect against attempts to override the agent’s instructions.
3. Secrets Masking: Toggle the Secrets switch ON to automatically detect and mask sensitive information like API keys and tokens.

​

Step 3.3: Configuring Data Privacy (PII Redaction and Keywords)

This is crucial for the Compliance Agent to handle sensitive HR data safely.

1. PII Redaction:
   • Activate: Toggle the Personally Identifiable Information (PII) switch ON.
   • Select Entities and Action: For a Compliance Agent, set the action to Redacted for most items:
     • Names (person), Email Addresses, Phone Numbers, SU Social Security Numbers (SSN): Select Redacted.
     • Credit Card Numbers: Select Blocked (as this is high-risk financial data).
2. Keywords (Secrets Masking):
   • Activate: Toggle the Keywords switch ON.
   • Use Case: Input proprietary or confidential terms (e.g., ProjectX, Q4_Forecast) into the field and set the action to Blocked or Redacted.

​

4. Applying the Policy and Enabling Final Guardrails

After the policy is created, you must apply it to your agent and ensure the remaining Responsible AI features are activated on the agent level.

​

Step 4.1: Attaching the RAI Policy and Enabling Memory

1. Return to Agent Builder: Navigate back to the “Compliance Advisor Bot” configuration screen.
2. Enable Memory: Toggle the Memory switch ON under Core Features.
   • Configure Memory: Set the Memory Provider to Lyzr Memory and set the Max. messages stored as Short-term Memory to 10 messages for context.

3. Apply RAI Policy: Toggle the Responsible AI switch ON under Safe & Responsible AI (referencing r2.png).
   • Configure: Click the settings icon and select the HR_Compliance_Policy you created.

​

Step 4.2: Activating Fairness and Hallucination Guardrails

These features work alongside your policy and RAG to guarantee the highest standards of safety and accuracy.

1. Fairness & Bias: Toggle the Fairness & Bias switch ON (r2.png).
   • Context: This feature actively monitors and mitigates potential biases (e.g., gender, racial) in the agent’s responses, ensuring equitable policy interpretation.
2. Hallucination Manager: Toggle the Hallucination Manager switch ON (r2.png).
   • Context: This provides active, enhanced verification against your Knowledge Base, preventing the agent from guessing or inventing policy information.

​

Step 4.3: Final Deployment

1. Final Review: Verify the Name, Description, Model, Agent Role, and Agent Instructions are set. Ensure Knowledge Base, Memory, Responsible AI, Fairness & Bias, and Hallucination Manager are all active.
2. Deploy: Click the Create or Deploy button to launch your fully governed, compliant AI agent.

​

5. Visual Verification of Responsible AI Guardrails

RAI Feature Tested: Personally Identifiable Information (PII) Redaction and Policy Refusal. Scenario: The user attempts to ask a policy question while including sensitive personal data (Name and SSN) in the input.

​

Test 5.2: Secrets Masking and Confidentiality Test

RAI Feature Tested: Secrets Masking (Keywords) and Prompt Injection Protection. Scenario: The user asks a policy question but includes a secret/confidential keyword (ProjectX) that was flagged for blocking in the RAI Policy (referencing Section 3.3).

​

Test 5.3: Policy Conflict Test (External PII Reference)

RAI Feature Tested: External PII Detection and Policy Refusal on Data Leakage. Scenario: The user asks a question about a policy but provides an external piece of PII (an email address) in the prompt, which is flagged for Redaction/Blocking in the policy.

​

Conclusion

The successful execution of these tests, as confirmed by the visual outputs, verifies that the custom HR_Compliance_Policy is enforcing:

• Privacy: Blocking or redacting PII to meet compliance standards (Test 5.1).
• Confidentiality: Preventing policy discussion when internal secrets are included in the prompt (Test 5.2).
• Security: Successfully identifying and blocking inputs that pose a security risk or violate the privacy mandate (Test 5.3).

The “Compliance Advisor Bot” is now fully vetted and ready for deployment.