The developers created an AI assistant for e-commerce sites that answers customer queries based on specific website documentation. The system uses a Python-based architecture to distinguish between casual conversation and factual information, ensuring that the AI does not hallucinate answers regarding company policies, delivery, or pricing.
To prevent misinformation, the model is restricted to providing facts only when it can cite a specific source from the site's knowledge base. The system uses a JSON schema where the 'source' field must contain the index of the retrieved documentation. If no relevant information is found, the bot is instructed to admit it does not know the answer rather than guessing.
The assistant integrates a fallback mechanism that offers to connect a human operator when the AI cannot provide a factual answer. To avoid false positives, the team implemented a rule-based system using regular expressions that handles common intents, such as requests for human support or order status, before the LLM is even invoked.
The developers use a two-part testing suite to validate the system: a set of 84 custom phrases related to e-commerce and 386 phrases from the Amazon MASSIVE dataset. This testing ensures that the bot correctly identifies user intent and avoids triggering human intervention for irrelevant queries, such as product names containing keywords like 'manager'.
Unanswered questions are collected in a 'Gaps' section of the owner's dashboard, allowing site owners to add new documentation or manual corrections. These corrections receive a priority bonus during the retrieval process, and the system allows for real-time updates to the assistant's behavior without requiring model retraining.
