Skip to main content

AI Knowledge Base Agent

The AI Knowledge Base Agent enables you to create an autonomous AI agent trained specifically on your company’s proprietary data. By uploading documents, scraping website URLs, and automatically indexing historical WhatsApp chats, the AI learns your organization’s context to handle customer inquiries intelligently.

Overview

The Knowledge Base Agent combines multiple data sources into a unified vector database (powered by Qdrant), allowing your AI to provide accurate, context-aware responses based on your organization’s knowledge.
All vectorized data is strictly partitioned using tenant/organization IDs to ensure complete data isolation between different Eazybe clients.

Key Features

Multi-Source Knowledge Base

Train your agent using three powerful data sources:
1

Web Links

Input any URL to scrape website content, help center articles, documentation, or blog posts. The agent automatically extracts and indexes the text content.
2

File Upload

Drag and drop documents directly into the platform. Supported formats include:
  • PDF files
  • DOCX documents
  • TXT files
  • CSV spreadsheets
3

WhatsApp Chat Sync

Toggle the Sync WhatsApp History option to automatically fetch and vectorize historical organizational WhatsApp chats. This provides the AI with real conversational context and previously resolved customer queries.

Agent Configuration

Customize your agent’s behavior with these settings:
  • Agent Name: A unique identifier for your agent (e.g., “Acme Support Bot”)
  • Role/Persona: Define how the AI should behave (e.g., “You are a helpful customer support representative for Acme Corp”)
  • Base Prompt: Set the foundational instructions that guide all agent responses
  • Fallback Action: Configure what happens when the AI is uncertain (e.g., “Handover to human agent”)

Real-Time Indexing Progress

Vectorizing large documents or thousands of chats takes time. The interface provides clear visibility into the indexing process:
Progress: 75% Indexed - 150/200 pages processed
Do not close the browser during initial indexing. However, once indexing begins, you can navigate away—the process continues asynchronously in the background.

One-Click WABA Deployment

Once your agent is fully trained and indexed, deploy it to any connected WhatsApp Business API (WABA) number:
1

Navigate to Deploy

Go to the Deploy tab within your agent configuration
2

Select WABA Number

Choose from a dropdown of your connected WhatsApp Business numbers
3

Deploy

Click Deploy to WhatsApp to activate the agent

How It Works

1. Creation Phase

Click Create New Agent and provide a name and description for your AI agent.

2. Data Ingestion Phase

Upload your knowledge base materials:
  • Drag and drop product PDFs, policy documents, or training materials
  • Paste URLs from your website, help center, or documentation
  • Enable Sync WhatsApp History to include historical conversations

3. Processing Phase

The backend processes your data through these steps:
  1. Chunking: Large documents are broken into manageable segments
  2. Embedding Generation: Each chunk is converted to a vector embedding
  3. Storage: Embeddings are stored in Qdrant with proper tenant isolation
  4. Indexing: A searchable index is built for fast retrieval
During this phase, you’ll see a real-time progress indicator showing the percentage complete and number of items processed.

4. Ready State

Once indexing reaches 100%, the agent’s status changes to Ready to Deploy. You can review the indexed sources and make adjustments if needed.

5. Deployment Phase

Select your WABA number and deploy. The AI now intercepts incoming messages on that number and responds based solely on your ingested knowledge base.

6. Live Operation

The deployed agent handles customer inquiries 24/7, using only the information from your knowledge base to generate accurate, contextually relevant responses.

Important Considerations

Data Isolation & Security

All vectorized data (especially WhatsApp chats) pushed to Qdrant is strictly partitioned using tenant/organization IDs. This prevents data leakage between different Eazybe clients.

Long-Running Uploads

For large knowledge bases (e.g., 5,000+ WhatsApp chats):
  • Indexing happens asynchronously via a background worker
  • The UI polls for status updates every few seconds
  • You can navigate away from the page during indexing
  • Return anytime to check progress via the agent dashboard

Stale Knowledge

If your knowledge sources change (website updates, new product docs, policy changes):
  1. Navigate to your agent’s Data Sources section
  2. Click Re-sync next to the affected source
  3. The system updates the embeddings in Qdrant with the latest content

Fallback Routing

When the AI encounters a query it cannot confidently answer based on the knowledge base, it follows your configured fallback action:
Recommended fallback actions include:
  • Handover to human agent - Transfer to a live team member
  • Collect information - Gather customer details for follow-up
  • Provide generic response - Offer a standard reply with contact options

Best Practices

Start Small

Begin with a focused knowledge base for a specific use case (e.g., product FAQs, return policy). Expand gradually as you monitor performance.

Quality Over Quantity

  • Use well-structured, up-to-date documents
  • Clean PDFs with clear formatting produce better results
  • Avoid duplicate or conflicting information

Monitor and Iterate

  • Review conversation logs to identify knowledge gaps
  • Add missing documents to improve coverage
  • Refine the base prompt based on real interactions

Test Before Deploying

Always test your agent thoroughly with sample queries before connecting it to your live WABA number.

Troubleshooting

Agent Not Responding

  • Verify the WABA number is properly connected
  • Check that the agent status is Live (not just Ready)
  • Ensure the knowledge base indexing completed successfully

Inaccurate Responses

  • Review the knowledge base for conflicting information
  • Check that relevant documents were properly indexed
  • Refine the base prompt to be more specific

Indexing Stuck

  • Large file uploads may take considerable time
  • Check file sizes—files over 50MB may need to be split
  • If stuck for more than 30 minutes, contact support

Missing WhatsApp Chats

  • Verify the Sync WhatsApp History toggle is enabled
  • Check that chat permissions are granted
  • Ensure chats are within the sync date range

FAQ