Course model

The Learning Companion uses a flexible model with feature flags. It supports

• A timeline-based view on course slots with concrete dates, and allows for associating directly with the slot (like recordings or boards), or indirectly via course sections.
• A topic-based view on all the provided materials and course slots.
• A structural view putting the whole course into a more traditional table of contents
• Conversational view using a state-of-the-art LLM-based chatbot

Content management

• Easy to use admin backend
• Extensive use of convention over configuration to reduce the
• Flexible and extensible use of scalable object storage as a native store

Chatbot features

• Based on GPT-4o as a standard model, the model to be used can be configured (for each of the courses of components separately)
• Querying indexed documents (with metadata), implementing a RAG approach with per-course configuration of major parameters (like number of chunks)
• Token streaming to the UI for low-latency answers
• Configurable history management
• Generate topic suggestions
• Generate follow-up question suggestions
• Inline citations of the used materials for better traceability (with direct quotes)
• Multiple and configurable (in CMS) prompts chosen at runtime (dynamic prompt selection) (either via LLM or direct similarity search)
• Sidebar with suggested course materials relevant to the question
• Thumbs voting to improve the quality
• Extensive chatbot interaction logging and self-evaluation via LLM-as-a-judge
• Generate clarifying info prompts to the user
• Specialized tools (or functions) using an agentic workflow to make the LLM automatically choose what to include - especially useful for API-based integration of content such as timetables, staff directories, meal plans, etc.
• Supports advanced UML model generation via PlantUML rendering and specialized prompts
• Low-barrier improvement via directly supported Q&A

Flexible document indexing

• Wide range of documents such as pdf, md, docx, pptx, but also video transcripts (both vtt and YouTube)
• Periodic indexing of web sites and other content
• Observability of workflow execution
• Multiple indexing methods, including multimodal indexing
• Configurable indexing parameters, such as chunking strategy
• Easy reindexing and chunk inspection from the CMS

Evaluation and quality monitoring

• Anonymous logging
• Self evaluation using LLM-as-a-judge approaches
• Manual quality tagging and annotation for improvement
• Auto-expiring information (e.g., semester-related)

Extensible and customizable

• Many feature flags, supporting, e.g., synchronous courses (with timeline), asynchronous courses, or course-indepenten bots
• Themeable, supporting automatic and manual switching between light and dark mode
• Easy addition of new material types
• Supports multiple languages (or a course level)
• Models, prompts, and chatbot parameters directly configurable in CMS
• Extensive API and views designed to be embedded into other sites (fragments and OEmbed)
• Planned: compose chatbots from multiple courses or more specialized bots

Modern technologies

• FastAPI with Jinja2 (with JinjaX components) and extensive use of Pydantic
• Request handling and service invocation completely asynchronous, maximizing throughput
• Hypermedia-first approach using HTMX
• Responsive web design
• Built ontop of langchain
• Prefect as orchestrator for indexing tasks
• Pinecone Serverless as vector database
• MongoDB Atlas for history management
• Log observability with Pydantic's LogFire, GenAI observability with LangFuse