Training AI

Overview

Zaapi’s AI Agent can handle customer messages, allowing your team to focus on inquiries that require human attention. To ensure accurate and helpful responses, training the AI is essential. The AI uses Knowledge Sources and Scenarios to generate responses:

• Knowledge Sources – Reference materials such as FAQ documentation, product descriptions, or help center articles that the AI uses to answer customer inquiries.

• Scenarios – Step-by-step instructions for handling specific customer situations in a predictable and consistent manner. These can also define when the AI should escalate an issue to a human agent.

• Quick Replies – References saved quick replies based on keywords to provide responses to customers (this can be turned off in the "Test AI" area)

Regularly updating the training documentation helps improve AI accuracy. If the AI cannot answer a customer’s question, be sure to add it to the training documentation.

How to structure training documentation

Structure Content with AI Agents in Mind

When building a knowledge base for AI, the goal is to make it easy for the model to understand, categorize, and retrieve relevant information. This means optimizing your content’s structure, formatting, and language to align with how AI ingests and interprets data.

Use Clear, Consistent Headings

AI relies heavily on headings to understand the structure and context of your knowledge base.

• Use proper HTML heading tags (<h1>, <h2>, etc.). Avoid using bold or large text for headings—AI may not recognize these as structural markers.

• Every heading should accurately describe the content beneath it. Avoid generic or misleading titles.

• Group related content together, and avoid overlapping topics under a single heading.

The AI agent uses these headings as context when chunking content. The more accurate and descriptive your headings, the more likely the AI will surface relevant information during a conversation.

Organize from General to Specific

Structure your content hierarchically:

• Start with broad categories, then narrow down to specific actions or scenarios.

• This helps the AI understand relationships between topics and retrieve the right level of detail based on the question.

Use Task-Based, Action-Oriented Language

AI performs better when articles are clearly tied to specific user intents.

• Use verbs in titles (e.g., Connect your Instagram account, Update business hours).

• Front-load keywords—put the most important words at the beginning of the heading or sentence.

This improves match relevance between user questions and content.

Avoid Assumed Knowledge

LLMs don’t inherently understand dependencies between articles unless context is made explicit.

• Don’t assume the AI has seen previous content. Link to prerequisite articles or include a brief summary if needed.

• Avoid internal jargon or product-specific terms without explanation, especially early in an article or heading.

One Topic per Section

Keep each chunk of content focused.

• Avoid combining multiple topics in one section. AI works best when each chunk maps to a single intent or question.

• Split complex workflows into separate, clearly labeled steps or articles.

---

How to train your AI

Using Knowledge Sources

1. Access the AI Agent Section – Go to the AI Agent section in the side navigation.

2. Click on "Knowledge Sources" – This is where you manage AI reference materials.

3. Download a Training Template – Choose the General Q&A Template (download) or the Product Details Template (download).
4. Fill in the Template – Add all relevant information, ensuring correct heading sizes, no empty cells, and no images (AI cannot process images).

5. Upload the Knowledge Source – Click "Add Knowledge Source" and give it a recognizable name (e.g., “FAQs”).

6. Wait for Upload Completion – This may take up to 20 minutes.

   • Successful – The entire knowledge source was uploaded.

   • Partial – Only part of the content was added due to character limits.

   • Failed – The upload was unsuccessful. Review the document format and try again.

Using Scenarios

1. Go to the AI Agent Section – Navigate to the Scenario Training section.

2. Click "Add Scenario" – This allows you to create a new AI response pattern.

3. Define the Scenario Title – Example: "If customer asks for a refund."

4. Add Trigger Phrases – List variations of customer inquiries that should trigger this scenario, such as:

   • "I want a refund"

   • "Can I get my money back?"

   • "My order didn’t arrive"

5. Set the AI Response – Choose between:

   • Following a set of instructions (clear step-by-step guidance).

   • Escalating to a human agent immediately.

6. Provide Clear Instructions – Example for a refund request:

   1. Ask the customer when they ordered the product.

   2. Inform them that if the order is over two weeks old, refunds are not possible.

   3. If under two weeks, request a video of the unboxing to verify product condition.

7. Click "Create Scenario" – The AI will now use this scenario when responding to similar inquiries.

---

AI Training Limitations

• Maximum of 1,000,000 characters for training.

• PDF files cannot be uploaded.

• Marketplace websites (Shopee, Lazada, TikTok Shop) cannot be scraped as a knowledge source.

• The AI does not differentiate between brands based on the channel (improvements coming soon!).

• AI does not use images in responses.

Regularly refining training content ensures the AI delivers accurate and effective customer support.