AI Assistant Best Practices

To get the most out of your WhatsCloud AI Assistant, follow these best practices for crafting prompts, preparing documents, and managing the user experience.

1. Crafting Effective Prompts (System Instructions)

The Prompt is the most important setting. It defines the AI's personality, rules, and boundaries. Be specific and clear.

Good Prompt (Specific & Role-Based):

"You are 'WhatsCloud Bot', a friendly and professional support assistant for OWNcodes. Your goal is to answer customer questions based *only* on the provided knowledge base. If the answer is not in the documents, politely say, 'I'm sorry, I don't have that information, but I can connect you with a human agent.' Do not make up answers."

Bad Prompt (Vague):

"answer questions"

Tips for Prompts:

• Define a Persona: Give it a name ("WhatsCloud Bot") and a role ("friendly and professional support assistant").
• Set Knowledge Boundaries: Use a phrase like "based *only* on the provided knowledge base." This prevents the AI from "hallucinating" or using its general knowledge.
• Define Fallback Behavior: Tell it *exactly* what to say when it doesn't know the answer (e.g., "I'm sorry, I don't have that information...").

2. Preparing Your Knowledge Base

The AI is only as smart as the documents you give it. Garbage in, garbage out.

• Quality over Quantity: A 10-page, well-structured PDF with clear answers is better than a 500-page unorganized document dump.
• Clear Headings: Use clear, descriptive headings in your .docx or .pdf files (e.g., "How to Get an API Key," "Pricing Plans"). The AI uses these headings to find relevant information.
• Simple Language: Write your documents in simple, clear language. Avoid complex jargon if possible.
• Factual Data: Ensure your documents contain facts, not long, rambling opinions. Q&A-style documents work very well.
• Small & Focused: It's better to upload 5 separate PDFs (e.g., Pricing.pdf, Features.pdf, API.pdf) than one giant 100-page "Everything.pdf".

3. Training and Testing

• Retrain After Uploads: Any time you upload, delete, or change a file, you must click the "Train Assistant" button.
• Wait for Sync: After training, wait for the status to show "Fully Synced" before testing. This can take a few minutes.
• Test Extensively: Ask questions you know are in the documents. Then, ask questions you know are *not* in the documents to test your fallback prompt.

4. User Experience (UX)

• Set Expectations: Use the "Message Footer" setting to add a line like "This response was generated by AI." This manages user expectations.
• Use a Delay: Set the "Message Delay (Seconds)" to 2 or 3. An instant answer can feel robotic. A short delay makes it feel more natural, as if someone is typing.
• Provide an "Escape Hatch": Always provide "Stop Keywords" (e.g., stop, agent, human). This gives users a clear way to exit the AI conversation and reach a person.

5. Maintenance and Monitoring

• Review Chats: Periodically review the AI's conversations with users (in the main Chat window) to see where it's failing or giving bad answers.
• Update Documents: If you see the AI consistently failing on a specific topic, update or add a new document to its knowledge base with the correct information and retrain it.

6. Cost Management

Using AI costs money, and API calls can add up. Here is how you control costs:

• Model Selection:
  • Use gpt-3.5-turbo or gpt-4o-mini for most standard Q&A bots. They are fast and very low-cost.
  • Only use expensive models like gpt-4 if you have a very complex reasoning task that the smaller models fail to handle.
• Monitor Your OpenAI Dashboard: Log in to your OpenAI account regularly to check your API usage and spending.
• Set Billing Alerts: In your OpenAI account, set up billing alerts to get an email if your usage exceeds a certain budget.

---