Quizgecko New Tech Ideas.docx

### Idea 1: Implementing Retrieval Augmented Generation for Quiz Generation [What is RAG? ] RAG incorporates retrieval mechanisms to provide more accurate and specific responses. It allows LLMs to access and utilize additional knowledge beyond their initial training data, enabling them to answer m...

### Idea 1: Implementing Retrieval Augmented Generation for Quiz Generation [What is RAG? ] RAG incorporates retrieval mechanisms to provide more accurate and specific responses. It allows LLMs to access and utilize additional knowledge beyond their initial training data, enabling them to answer more specific questions based on provided information. You take data you want to use as sources, you label this data, you then embed this data, then when a user enters a query that outputs the most similar sources, which are then fed into the LLM for generation. [How can we use it? ] 1. LABEL: Label and Chunk stored documents and uploads. 2. EMBED: Embed these documents with links back to the full source. 3. QUERY: When a user enters a quiz prompt, relevant document ids are outputted. 4. CONTEXT: The document ids are then used to fetch the full documents, which are retrieved context is combined with the user\'s question. 5. GENERATE: A language model uses the question and retrieved context to generate a quiz. We could also extract papers from pubmed and embed all of these, not just the users uploads, then when we generate quiz we can reference scientific studies that were used in generation. [What are the benefits? ] 1. More accurate and specific quizzes as its based on users uploads. 2. Users can see sources of what was used to generate that quiz. 3. Greate for using AI to generate specific quizzes for SEO (e.g. biology AQA GCSE). ### Idea 2: Taking Idea 1 a step further: Integrating a Knowledge Graph A diagram of a family tree Description automatically generated [What is a knowledge graph? ] 1. Nodes: Representing entities (e.g., topics, concepts, documents) 2. Edges: Representing relationships between entities 3. Properties: Additional information about nodes and edges [What would the relationships be? ] 1. Document-Topic: Link documents to relevant topics 2. Topic-Subtopic: Create hierarchical relationships between topics 3. Document-Concept: Connect documents to the concepts they contain 4. Concept-Concept: Link related concepts [How can we use it? ] We can put all of the data we have gathered so far into the knowledge graph with very predefined categories (e.g. Biology would have child GCSE, which would have child AQA or OCR). By mapping this data into these different categories it will allow us to rapidly and accurately generate very specific quizzes right down to the exam board. Labelling the data would be the hardest thing, but if we can crack that we can unlock so much value. [What are the benefits?] 1. When we generate a quiz we can rapidly find relevant uploads and material. 2. Allowing quizzes to be more accurate and the user can see the sources used to generate a quiz. 3. Users can see sources of what was used to generate that quiz. ### Idea 3:Agentic AI for course specific chatbots and feedback [What is Agentic AI] - It is a GPT than can prompt itself until a goal is met. - You define the goal and let it run until that goal is met. [How can we use it?] 1. Homework helper to respond to more complex university level queries. [What are the benefits?] 1. Develop AI agents specializing in particular subjects to assist students with their studies. 2. Create AI agents that can evaluate and provide feedback on homework assignments or quizzes submitted by students. 3. Implement AI agents that continuously analyze and enhance the quality of generated quizzes over time. ### Idea 4: Small Language Models [What is a small language model?] - A SLM) is a version of a language model with much fewer parameters than LLMs. - Typically has up to 1 billion parameters, compared to billions or trillions in LLMs - More streamlined architecture and simpler design - Requires less computational power and training data - Faster to train and deploy, often taking minutes or hours instead of days or weeks [How can we use it?] 1. Localise SLMs for basic quiz generation offline on people devices. 2. Using information from the knowledge graph you could have SLMs for each course that generate highly fine tuned and accurate quizzes. [What are the benefits?] 1. Way higher course specific quiz quality = we beat the competition. 2. Offline AI quiz generation -- no competitor has this. ### Idea 5: Latex Based Past Paper Generator [What is a latex based past paper generator?] - We build out a fine tuned LLM trained on all the past papers we have uploaded. - It will then be able to generate past paper styles quizzes with much more complicated questions that look and feel like a real past paper. - Integrate this with RAG and the KG we have a very powerful feature that will be best in class. [How can we use it?] 1. To generate course specific past papers that look and feel like university level past papers or A-level/GCSE past papers. [What are the benefits?] 1. We will be the only AI study app that has the ability to do this. ### Idea 6: Explainable AI integrated into the homework helper [What is a explainable AI?] - An AI similar to o1 that can explain its reasoning process. - Leading to less hallucination and more accurate responses. [How can we use it?] - We can use it in the homework helper so the student can see the thought process behind the AI's answer which will increase the amount they learn. [What are the benefits?] - Again we will be the only AI study app incorporating this technology. ### Idea 7: Videos and Songs with GPT-4o Voice