System User Guide
Knowledge Text Pre-processing
After retrieving each article’s title, body, and URL, the content is stored as plain text in a MongoDB article database. To support RAG-based article retrieval, the title and body are merged into an index text, which is then converted into a fixed-length vector using Sentence-BERT’s paraphrase-multilingual-mpnet-base-v2 multilingual model and stored in the article vector database.
Module Construction
Search Architecture
A resource recommendation algorithm combiningBM25 (term frequency), TF-IDF, co-occurrence analysis, and SBERT embedding technology to search science education resources based on user-entered keywords.Response Architecture
Built onRAG and chain-of-thought prompt design, using a large language model (GPT-4 Turbo) to generate appropriate responses and follow-up questions.
Trello Data Creation Flow
Prompt List
| role | prompt |
|---|---|
| system | #zh_tw You are a natural science middle school teacher who uses Traditional Chinese and is speaking with a student who is new to research. The student is researching Research Topic. You need to help the student understand related knowledge, so you should generate a response and ask the student a follow-up question to promote deeper thinking about the Research Topic. The output text must be based on Markdown syntax. |
| user | *chat text (oldest to newest)* |
Response Module Data Pre-processing
Chat Text Processing
Knowledge Text Processing
Search Module
Building a Self-Directed Learning Resource Recommendation System Using Natural Language Processing
A resource recommendation algorithm combining BM25 (term frequency), TF-IDF, and SBERT embedding technology to search science education resources based on user-entered keywords.