RAG Teaching Material Retrieval-Augmented
Methodology

1. Overview

The AI teaching assistants (UeduGPTs) on the Uedu platform support the AI knowledge base function: Instructors upload Course materials (PDF, DOCX, PPTX), and the system automatically chunks and vectorises the materials. When students ask questions, it performs semantic retrieval and injects the most relevant material excerpts into the AI's System Prompt so that the AI's replies can be grounded in the Course materials.

This mechanism is known as RAG (Retrieval-Augmented Generation), which is the current mainstream method used in the industry to enable LLMs to answer with evidence. This document explains the implementation details of Uedu's RAG for researchers to understand the data generation process.

2. What is RAG

RAG (Retrieval-Augmented Generation) is an architecture that combines information retrieval and text generation (Lewis et al., 2020). Its core concept is:

1. Retrieval: when the user asks a question, first find the most relevant document excerpts from the knowledge base
2. Augmentation: inject retrieved passages into the LLM input as additional context
3. Generation: LLM generates answers based on the original question + retrieved teaching materials

Compared with answers that rely only on an LLM’s built-in knowledge, RAG allows AI replies to be grounded in specific teaching materials, reducing the risk of hallucination and ensuring the content remains relevant to the Course.

3. Data processing workflow

After the Teacher uploads course materials, the system carries out the following processing steps in the background:

3.1 Text extraction

The system supports three file formats, and uses the corresponding Python packages to extract the full text:

3.2 Semantic chunking (Chunking)

The extracted full text is tokenised into chunks at token level using the tiktoken cl100k_base tokenizer:

• Maximum chunk length: 800 tokens
• Adjacent chunk overlap: 100 tokens — ensure the semantics across chunk boundaries do not break
• Minimum chunk length: 20 tokens — chunks that are too short will be discarded

Chunking is carried out page by page (or slide by slide). If the text on a single page is no more than 800 tokens, treat the page as one whole chunk; otherwise, use token-level sliding segmentation.

3.3 Vector embedding (Embedding)

The text of each chunk is converted by OpenAI text-embedding-3-small into a 1536-dimensional dense vector, stored in the database BLOB field as float32.

4. Semantic retrieval mechanism

When a student submits a question, the system carries out the following retrieval steps before replying:

1. Convert the Student's question text into a 1,536-dimensional vector using the same embedding model (text-embedding-3-small)
2. Load the embedding matrix for all chunks in this Course from memory cache (cache TTL 5 minutes)
3. Compute cosine similarity between the question vector and all chunk vectors
4. Retrieve the top-k chunks with the highest similarity, filtering out results below the threshold
5. Format the retrieved teaching material excerpts and inject them into the AI's System Prompt

5. System parameters

6. GraphRAG expansion

Uedu supports advanced GraphRAG mode. When the Instructor turns on this mode, the system will not only use standard vector retrieval, but will also useKnowledge graphProceedGraph expansion retrieval：

1. Standard vector retrieval hits several chunks
2. Query the knowledge concepts corresponding to these chunks (kg_concepts)
3. Use the relation edges in the knowledge graph (1-hop) to find adjacent concepts
4. Supplement the retrieval results with teaching material chunks from adjacent concepts (up to 3 extra)

When expanding the graph, the edge weights for prerequisite (prior knowledge) and contains (containment relationship) will receive a 1.2x boost, and teaching materials for these two types of links will be prioritised.

7. Retrieval logs and transparency

Detailed logs for each retrieval are stored in the rag_retrieval_log table, including:

• query_text: the Student's original question
• chunk_ids: array of chunk IDs matched by retrieval
• similarity_scores: corresponding cosine similarity scores
• retrieval_time_ms: retrieval time (milliseconds)

These records allow researchers to analyse “which textbook segments AI referred to when answering” and “the degree of semantic match between Student questions and the teaching materials”.

8. Suggested research citation

It is recommended to add the following information as a note as well:

• The Course uses standard RAG or GraphRAG mode
• Number of uploaded course materials and total token count (viewable from the Instructor side)
• embedding model name and version