website-assistant/AI Fallback Strategy

7.10 AI Fallback Strategy

When a visitor asks a question in the Website Assistant, Pop first attempts to retrieve answers from the Knowledge Base (KB). If the KB does not return a match, the snippet is not relevant enough, or confidence is too low, the system automatically enters the AI Fallback Strategy.

Fallback ensures that visitors always receive a helpful response and never experience a "dead end" during a conversation.

Pop supports three fallback modes with configurable priority rules.


🧩 1. When Fallback Is Triggered

Fallback activates under the following conditions:

Condition Description
KB Miss No retrieval results
Low Relevance Similarity score falls below threshold
Invalid Content Snippet too short, incomplete, or corrupted
KB Disabled This site has KB integration turned off
Unsuitable for KB Open‑ended questions, chit‑chat, subjective asks

When any of these occur, fallback is triggered automatically.


🧩 2. Fallback Modes (Configurable)

You can choose one of the following three modes in the admin panel.

Mode 1: Fixed Reply

The system returns a preconfigured static message.

Suitable for:

  • No KB prepared yet
  • Strong content safety requirements
  • Scenarios where AI answers must be restricted

Example:

Sorry, I’m unable to answer this question right now. Please contact the administrator or check the Help Center.

Mode 2: AI Response

The system sends the visitor’s question to the AI model and returns the generated response.

Suitable for:

  • KB covers only part of the domain
  • AI is stable and helpful
  • You want intelligent fallback when KB misses

AI uses:

  • The site’s current model configuration
  • Assigned persona (if enabled)
  • Conversation history (if enabled)

Mode 3: Hybrid Priority Strategy

A multi‑layer fallback combining both fixed reply and AI.

Example strategies:

Fixed Reply → AI (official recommended: safer)
AI → Fixed Reply (more intelligent while ensuring a final safe fallback)

Available priority modes:

Priority Mode Description
Fixed → AI Safe and stable (recommended)
AI → Fixed More intelligent
Fixed Only Maximum control
AI Only For fully AI‑driven sites

🧠 3. Internal Fallback Logic Flow

The complete flow:

User Question
   ↓
Knowledge Base Retrieval (Hybrid)
   ↓
Matched?
   ├─ Yes → Return KB Answer (with snippet)
   └─ No → Enter Fallback

Fallback:
   ├─ Fixed Reply → Return static message
   ├─ AI Response → Generate via model
   └─ Hybrid → Try based on configured priority

This guarantees a response in all scenarios.


⚙ 4. Admin Configuration Options

In Website Assistant → Support Settings, you can configure:

Setting Description
Enable Knowledge Base Fallback triggers only when KB misses
Fallback Mode Fixed / AI / Hybrid
Fixed Reply Content Text returned when fixed mode is used
AI Model Model used during fallback
AI Temperature Controls creativity level
Similarity Threshold Determines what counts as KB “match”
Risk Word Filtering Prevents AI from outputting sensitive content

🧪 5. Examples

Example 1: "What is your pricing?"

  • KB has no pricing document
  • Fallback mode: Fixed Reply

System returns:

“Pricing information is currently unavailable. Please contact the administrator.”


Example 2: "How do I call this API?"

  • KB matches the API documentation → snippet returned
  • Fallback does not activate

Example 3: "What do you think about the future of AI?"

  • KB cannot match
  • Fallback mode: AI Response

AI generates an open‑ended answer.


📝 6. Best Practices

✔ Use Hybrid KB Retrieval to reduce fallback frequency

A more accurate KB means a more professional assistant.

✔ Keep fixed replies concise

Clear and direct is best.

✔ For regulated industries, use: “Fixed Reply → AI”

Ensures safety and compliance.

✔ Monitor KB hit rate regularly

Low hit rate indicates KB improvement is needed.


📌 Summary

The purpose of AI fallback is not to replace the Knowledge Base. Instead, it ensures:

  • Visitors never receive a blank response
  • A consistent and smooth support experience
  • Controlled safety when content is uncertain

With proper configuration, you can balance stability and intelligence to fit your business requirements.