Customer support · Production

How Wix's AI Site-Chat Redefines Chatbots with Adaptive Feedback and Dynamic Knowledge

The problem

Traditional AI chatbots built on LLMs cannot customize responses based on business owner preferences, offer no feedback mechanism for owners to teach the chatbot, and rely on static knowledge bases and fixed system prompts that do not adapt to unwritten domain knowledge.

First attempt

Traditional RAG systems have static knowledge bases and fixed prompts that cause outdated responses and user frustration, while LLMs used directly for classification suffer from overconfidence, hallucination, prompt fatigue, and high inference cost.

Workflow diagram · grounded in source
1
Owner queries chatbot
trigger
“The owner asks a question (query) and receives a response from the chatbot.”
2
RAG answer generation
ai_action
“We utilize the Retrieval Augmented Generation (RAG) mechanism to produce an answer. This mechanism selects the most relevant content from the site to include in the context window.”
3
Owner provides feedback
human_review
“The owner provides feedback, which could be additional information, corrections, instructions, satisfaction or anything else.”
4
LLM feature extraction
ai_action
“utilize LLMs for feature extraction by answering a predefined set of yes/no/don't know questions about the input text, in parallel”
5
CatBoost classification
ai_action
“CatBoost Classifier uses the yes/no (and don't_know) responses as input features. CatBoost prepares the categorical data and helps us avoid overfitting.”
6
Knowledge document creation
output
“Based on the classification, it generates a new document: Knowledge Document: Contains information to enrich the context for future queries. Prompt Instruction Document: Includes instructions for modifying the system prompts.”
7
Vector database update
integration
“The new document is stored in the knowledge database and its embedding is added to the vector database for future retrieval.”
8
Adaptive response generation
ai_action
“The LLM generates a response using the query, retrieved documents, and the (possibly modified) system prompt.”
Reported outcome

The DDKI-RAG system provides a more adaptive and dynamic approach to knowledge retrieval and prompt modification, allowing real-time learning and adaptability and leading to more accurate and context-aware chatbot responses.

Reported metrics
Chatbot accuracy and user experienceimproved accuracy, efficiency, and user experience
Chatbot response qualitymore accurate and context-aware chatbot responses
Reported stack
LLMsRAGCatBoostDDKI-RAG
Source
https://www.wix.engineering/post/how-wix-s-ai-site-chat-redefines-chatbots-with-adaptive-feedback-and-dynamic-knowledge
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The DDKI-RAG system provides a more adaptive and dynamic approach to knowledge retrieval and prompt modification, allowing real-time learning and adaptability and leading to more accurate and context-aware chatbot res…

What tools did this team use?

LLMs, RAG, CatBoost, DDKI-RAG.

What results were reported?

Chatbot accuracy and user experience: improved accuracy, efficiency, and user experience; Chatbot response quality: more accurate and context-aware chatbot responses (source-reported, not independently verified).

What failed first in this deployment?

Traditional RAG systems have static knowledge bases and fixed prompts that cause outdated responses and user frustration, while LLMs used directly for classification suffer from overconfidence, hallucination, prompt f…

How is this customer support AI workflow structured?

Owner queries chatbot → RAG answer generation → Owner provides feedback → LLM feature extraction → CatBoost classification → Knowledge document creation → Vector database update → Adaptive response generation.

Related customer support cases