Haruka Network Documentation
Getting Started
Getting Started

Smart Reply

After receiving a Crisp message, the system first attempts to match Keyword Replies. If no match is found, it proceeds to the AI Smart Reply.

Please Note:

  • Smart replies consume your account balance. Costs vary depending on your subscription plan.
  • The average cost is approximately ￥0.002 per text reply, making it an extremely cost-effective solution.

How the AI Context is Built:

The AI constructs its response based on the following data points:

  1. System Prompts: Your configured AI instructions.
  2. Knowledge Base: Relevant snippets retrieved from your custom data.
  3. Crisp Metadata: (Information provided by Crisp, not actively collected by us)

  • User Profile (Reported by your website or the user).

  • Geographic Location.

  • Operating System & Browser.

  • Additional Data (Custom data attributes reported by your site).

  1. User Message & Timestamp: The actual content and time of the query.

AI System Prompts

Effective prompts can significantly improve communication efficiency. Below is a recommended template:

As a support agent, please always reply in a gentle and patient manner, ensuring the user feels respected and understood. 
Please use plain text only; avoid using Markdown or complex formatting.
Listen actively to user issues, ask for details patiently, and provide clear, detailed answers or guidance.
Prioritize the provided Knowledge Base content for accuracy. If the information is not available, direct the user to check here: [Your URL]

In the first reply to a user, include the following text to guide them on using keywords: 
"Reply 'Menu' for automated help. Human agents may be unavailable momentarily; please wait or use our menu system and knowledge base."

AI Knowledge Base

You can convert your documentation into an Embedding Model. You only need to provide the "Title" and "Content." This process consumes a very small portion of your balance but greatly improves AI accuracy.

Summary

You can choose to use Prompts only without a Knowledge Base. However, if you use the Knowledge Base, you must also configure a Prompt so the AI understands its identity and how to use the retrieved information.

Comparison: Knowledge Base Embedding vs. Pure Prompting

Feature Knowledge Base (Embeddings) Pure Prompt Replies
Logic Vectorizes data. Retrieves relevant snippets first, then sends them to the AI with the prompt. Writes the entire knowledge base into the prompt. AI references the full context every time.
Cost Only consumes tokens for Question + Specific Snippets. More cost-effective for large datasets. Consumes tokens for the Entire Dataset per message. Costs increase significantly as the base grows.
Speed Extra retrieval step, but smaller context leads to faster model inference. No retrieval step, but a massive context can significantly slow down model processing.
Quality Responses are focused and relevant, staying strictly on topic based on matched data. Provides comprehensive info, but prone to redundancy, rambling, or "hallucinations."
ON THIS PAGE