SiteGPT

Automatically discovers and indexes website content by crawling specified domains, extracting text, metadata, and structure from HTML pages. Uses recursive link-following with configurable depth limits and robots.txt compliance to build a searchable knowledge base without manual content uploads. The indexed content becomes the foundation for all subsequent AI responses.

Converts user queries into vector embeddings and performs semantic similarity matching against the indexed website content, returning contextually relevant passages even when exact keyword matches don't exist. Uses embedding models (likely OpenAI or similar) to understand query intent and match it against pre-computed document embeddings stored in a vector database, enabling natural language search without keyword precision requirements.

Generates customer support responses by combining retrieved website content with LLM reasoning, using a prompt engineering pattern that instructs the model to answer only based on provided context and decline out-of-scope questions. The system passes ranked search results as context window input to the LLM, enabling responses that cite specific pages and maintain consistency with documented information while preventing hallucination.

Maintains conversation state across multiple user messages by storing and retrieving conversation history, enabling the chatbot to understand context and answer follow-up questions that reference previous exchanges. Uses session-based state management to track user identity, conversation thread, and context window, allowing the LLM to reference prior messages when generating responses while managing token limits.

Provides a JavaScript widget that can be embedded on any website to display the chatbot interface inline, handling iframe rendering, styling customization, and event communication between the host page and chatbot iframe. The widget uses postMessage API for cross-origin communication and includes configuration options for appearance, behavior, and integration with the host site's analytics or CRM systems.

Detects when conversations exceed chatbot capabilities and routes them to human support agents, using rule-based triggers (keywords, sentiment, escalation requests) or confidence thresholds to determine when human intervention is needed. Preserves conversation history and context when handing off, allowing agents to continue the conversation seamlessly without requiring customers to repeat information.

Collects metrics on chatbot usage, conversation quality, and customer satisfaction, providing dashboards showing conversation volume, resolution rates, common questions, and user feedback. Analyzes conversation patterns to identify gaps in indexed content, frequently escalated topics, and opportunities for chatbot improvement through data-driven insights rather than guesswork.

Automatically detects user language from input and responds in the same language, using language detection models and multilingual LLM capabilities to handle conversations in multiple languages without separate configuration per language. Indexed content is searched across all available language versions, and responses are generated in the user's detected language while maintaining consistency with source material.

Allows users to provide additional training data, examples, or instructions to customize chatbot behavior beyond indexed website content, using techniques like prompt engineering, example-based learning, or fine-tuning to adapt the chatbot to specific terminology, tone, or business rules. Custom training data is incorporated into the context window or used to adjust response generation without requiring full model retraining.

Exposes REST API endpoints allowing developers to send queries and receive responses programmatically, enabling integration with custom applications, workflows, or systems beyond the standard web widget. API includes authentication, rate limiting, and structured response formats (JSON) for easy integration with third-party tools, CRMs, or internal systems.