Enterprise Knowledge Base Multi-Agent System

Executive Summary

Enterprise knowledge is scattered across HR, Finance, and IT departments, making it difficult for employees to find information efficiently. This article demonstrates how to build a Multi-Agent enterprise knowledge base Q&A system using Tencent Cloud ADP.

Key Highlights:

• "Orchestrator Agent + three specialized sub-Agents" hierarchical architecture
• Domain-specific retrieval through knowledge base tagging
• Each sub-Agent focuses on its domain for more accurate answers
• Unified knowledge base maintenance reduces operational costs

1. Pain Points in Enterprise Knowledge Management

1.1 Challenges with Traditional Knowledge Bases

Enterprises commonly face these knowledge management challenges:

1.2 Why Multi-Agent?

A single Agent handling the entire company knowledge base faces two core problems:

Problem 1: Blurred Knowledge Boundaries

When employees ask "How do I request vacation?" and "How do I apply for a server?", a single Agent searching through massive documents often results in:

• Long retrieval times with mixed, less relevant results
• Inability to adjust response style for different domains
• Difficulty handling complex cross-department questions

Problem 2: Access Control Challenges

Some knowledge involves sensitive information (e.g., compensation policies, financial procedures) requiring fine-grained access control.

2. System Architecture Design

2.1 Overall Architecture

This system adopts a "1+3" Multi-Agent architecture:

                    User Question
                         ↓
                ┌────────────────┐
                │  Orchestrator  │
                │    Agent       │
                │ (Intent Router)│
                └────────┬───────┘
                         ↓
        ┌────────────────┼────────────────┐
        ↓                ↓                ↓
┌───────────────┐ ┌───────────────┐ ┌───────────────┐
│   HR Agent    │ │ Finance Agent │ │  IT Agent     │
│(HR tag+Others) │ │(Finance+Others)│ │ (IT tag+Others)│
└───────────────┘ └───────────────┘ └───────────────┘
        ↓                ↓                ↓
        └────────────────┼────────────────┘
                         ↓
                  Expert Response

2.2 Role Distribution

2.3 Knowledge Base Tagging System

All documents are stored in a unified knowledge base, categorized by tags:

3. Building on Tencent Cloud ADP

3.1 Create Application

Step 1: Create a New Application

Log in to Tencent Cloud ADP and create a new Multi-Agent application, then navigate to "Knowledge":

1. Click "Import"
2. Create and add "File tag"
3. Upload department documents sequentially

Step 2: Upload Documents with Tags

When uploading department documents, assign corresponding tags:

Enterprise Knowledge Base/
├── [HR] Company HR Management System and Issue Handling Procedures.docx
├── [Finance] Company Financial System and Issue Handling Procedures.docx
├── [IT] Common Issues IT Operations Manual.md
└── [Others] IntelliFuture Tech Co.docx

3.2 Create Sub-Agents

HR Support Assistant (Agent) Configuration Example:

Role Definition:
You are an enterprise HR expert, specializing in answering employee questions about personnel, attendance, benefits, and training.

Task Description:
1. Answer questions based on documents tagged "HR" and "Others" in the knowledge base
2. Provide accurate, professional answers citing specific policy clauses
3. For uncertain questions, suggest contacting the HR department

Response Style:
- Friendly and professional, like a senior HR colleague
- Numbers (e.g., leave days) must be accurately cited from documents
- Present complex procedures as step-by-step lists

Finance Support Assistant (Agent) Configuration Example:

Role Definition:
You are an enterprise financial advisor, specializing in answering employee questions about reimbursement, budgets, invoices, and payments.

Task Description:
1. Answer questions based on documents tagged "Finance" and "Others" in the knowledge base
2. Ensure accuracy for amounts, deadlines, and key information
3. For complex financial questions, suggest consulting the Finance department

Response Style:
- Precise and rigorous, focusing on numbers and procedures
- For reimbursement questions, explain required materials and approval process
- Provide form download links when necessary

IT Support Assistant (Agent) Configuration Example:

Role Definition:
You are an enterprise IT support specialist, specializing in answering employee questions about system usage, device requests, and troubleshooting.

Task Description:
1. Answer questions based on documents tagged "IT" and "Others" in the knowledge base
2. Provide clear step-by-step instructions, with screenshots when helpful
3. For urgent issues, provide IT hotline and ticket submission methods

Response Style:
- Concise and clear with well-defined steps
- Explain technical terms for non-technical employees
- Provide quick solutions for common issues

3.3 Configure the Corporate Intelligent Operations Assistant Main Agent (Orchestrator Agent)

The orchestrator Agent handles intent recognition and task routing:

Role Definition:
You are the dispatch center for the enterprise intelligent assistant, responsible for understanding employee questions and routing them to the appropriate specialized Agent.

Task Description:
1. Analyze the domain of user questions: HR, Finance, or IT
2. Forward questions to the corresponding specialized Agent
3. For multi-domain questions, route based on primary intent

Routing Rules:
- Personnel, attendance, benefits, training, onboarding/offboarding → HR Assistant
- Reimbursement, invoices, budgets, payments, financial approvals → Finance Assistant
- Systems, devices, network, permissions, troubleshooting → IT Assistant
- When uncertain, ask the user to clarify their needs

3.4 Select Agent Collaboration Mode

Configure the Agent handoff relationships in "Settings":

3.5 Debug and Publish

Debugging Checklist:

1. Test typical questions from each domain to verify routing accuracy
2. Confirm sub-Agents correctly retrieve documents with their assigned tags
3. Validate cross-domain question handling logic
4. Verify answer accuracy and professionalism

Test Case Examples:

4. Technical Implementation of Domain-Specific Retrieval

4.1 Tag Filtering Mechanism

In Tencent Cloud ADP, configure tag filters when setting up knowledge base retrieval for each sub-Agent:

HR Agent Retrieval Configuration:

• Knowledge Base: Enterprise Knowledge Base
• Tag Filter: HR OR Other

Finance Agent Retrieval Configuration:

• Knowledge Base: Enterprise Knowledge Base
• Tag Filter: Finance OR Other

IT Agent Retrieval Configuration:

• Knowledge Base: Enterprise Knowledge Base
• Tag Filter: IT OR Other

4.2 Why Can Every Agent Access the "Others" Tag?

The "Others" tag stores company-wide information such as:

• Company overview, organizational structure
• Office locations, contact information
• General procedures, shared resources

This information may be needed in any conversation, so all sub-Agents should have access.

5. Results and Optimization

5.1 Performance Comparison

After deploying this system, one enterprise saw significant improvements:

5.2 Continuous Optimization Recommendations

1. Regular Knowledge Base Updates

• Update documents promptly when policies change
• Remove outdated content to prevent misinformation
• Add new documents based on frequently asked questions

2. Analyze Conversation Logs

• Identify questions with inaccurate answers
• Discover gaps in knowledge base content
• Optimize Agent prompts

3. Expand Agent Capabilities

• Integrate enterprise system APIs (e.g., attendance, reimbursement systems)
• Support action-based tasks (e.g., submit tickets, check balances)
• Add new specialized Agents (e.g., Legal, Administration)

6. Frequently Asked Questions

Q1: What if a user question spans multiple domains?

The orchestrator Agent routes based on the primary intent. If multi-domain collaboration is truly needed, you can configure the orchestrator to determine whether additional domain information is required after receiving the first response.

Q2: How do you handle questions not covered in the knowledge base?

Configure fallback logic in each sub-Agent's prompt:

• Clearly inform the user that no relevant information is available
• Provide human contact channels (e.g., HR email, IT hotline)
• Log the question for future knowledge base updates

Q3: How do you ensure sensitive information security?

• Don't upload sensitive documents (e.g., salary details) to the knowledge base
• Control access scope through tag permissions
• Explicitly prohibit disclosure of sensitive information in Agent prompts

Q4: Does the knowledge base require retraining after updates?

Tencent Cloud ADP's knowledge base supports real-time updates. Documents become searchable immediately after upload without model retraining.

Q5: How do you evaluate system effectiveness?

Key metrics to monitor:

• Resolution rate (Did users get satisfactory answers?)
• Routing accuracy (Were questions sent to the correct Agent?)
• Answer accuracy (Do answers match knowledge base content?)
• User satisfaction (Collected through feedback)

7. Use Cases and Limitations

7.1 Suitable Scenarios

7.2 Not Suitable For

Conclusion

The Multi-Agent enterprise knowledge base Q&A system achieves precise retrieval and expert answers through the "orchestrator routing + sub-Agent specialization" architecture.

Compared to traditional single-Agent solutions, this architecture offers better expertise, scalability, and maintainability, making it particularly suitable for knowledge management in medium to large enterprises.

Start now and build your enterprise knowledge base Q&A system with Tencent Cloud ADP!

→ Start Free Trial

Appendix: Quick Start Resources

Recommended Resources

• Tencent Cloud ADP Official Documentation
• Low-Code Multi-Agent Guide
• Corporate Intelligent Operations Assistant