My contribution

Full-Stack Developer

I built the browser chat widget, implemented the n8n chatbot workflows, and integrated the assistant into Moodle. My work included connecting Pinecone retrieval to answer generation in the RAG version. I deployed the widget and tested the complete path from a trainee's message through n8n to the reply displayed in the LMS.

Support inside the learning platform

• A floating launcher opens a chat panel for questions about LMS navigation, course activities, account access, technical requirements, and support channels.
• Configurable bot name, theme color, screen position, welcome message, and input placeholder let the same widget fit different LMS installations.
• Typing feedback, automatic scrolling, Enter-to-send, and a full-screen mobile layout support everyday use. The floating launcher hides while mobile chat is open so it does not cover the conversation.

Moodle-to-n8n request flow

• The standalone JavaScript widget sends a JSON POST request to an n8n webhook with the message, application identifier, session identifier, current LMS hostname, and recent conversation context.
• The workflow validates input, selects a support topic, builds the answer context, calls the LLM, formats the result, and returns JSON for the widget to display.
• The browser interface and workflow remain separate: widget settings select the endpoint, while n8n coordinates the support logic and model call.

Pinecone RAG version

• The widget sends a trainee's question to a self-hosted n8n webhook, which coordinates retrieval and answer generation.
• Pinecone provides vector retrieval over curated FAQ content. The retrieved context is supplied to the LLM to help ground its answer in the support material.
• n8n returns the generated response to the Moodle widget. This retrieval-augmented generation flow is a separate implementation from the keyword-based manual and FAQ workflow described below.

Manual & FAQ matching version

• The reviewed workflow contains a curated trainee-manual topic map and FAQ answers covering platform features, learning activities, completion guidance, and common technical questions.
• A keyword classifier scores matching phrases, giving longer phrases more weight. Manual content wins when its score equals or exceeds the FAQ match; otherwise the strongest FAQ supplies the context.
• The selected reference and a short conversation window are passed to an n8n LLM chain for a concise answer. When no topic matches, the assistant is guided to offer general LMS help or refer course-specific questions to the Help Desk or instructor.

Relevant guidance & support boundaries

• Successful manual-based answers can include an Official guide link built from the current LMS hostname and the matched manual page, helping trainees continue to the relevant reference.
• Workflow guidance focuses the assistant on platform support and navigation, with predefined responses for requests to complete academic assessments or disclose internal instructions.
• The workflow rejects empty or overly long messages. Response handling provides a temporary-unavailability message with FAQ and support alternatives when the model call fails.

Conversation continuity

• The widget stores the session identifier and conversation in browser local storage, separated by application identifier, so a conversation can be restored after navigation or reload.
• A Clear action asks for confirmation before removing the saved conversation and restoring the welcome message.
• Only a recent portion of the conversation is sent with each request, and the workflow narrows that context further before building the model prompt.

Implementation & deployment

• Plain JavaScript, HTML, and CSS provide a standalone widget that injects its interface into the host page without requiring a frontend framework.
• The repository includes Vercel configuration for serving the widget script, while a self-hosted n8n webhook handles the chatbot backend.
• I deployed the browser widget and tested the end-to-end Moodle-to-n8n response flow. The widget repository also includes a regression check for the mobile launcher overlap fix.