Curriculum-Driven RAG Educational Q&A System: Using AI to Reduce Hallucinations and Enhance Learning Experience

A RAG-based educational Q&A system built on NCERT textbooks, which effectively reduces hallucinations in large language models through FAISS vector retrieval, confidence filtering, and keyword verification mechanisms, providing students with more reliable learning support. Developed by Pruthviraj Khot from Pimpri Chinchwad College of Engineering in India, this system uses authoritative textbooks as knowledge sources and ensures answer accuracy through multi-layer mechanisms.

Large Language Models (LLMs) are widely used in education, but the hallucination problem (confidently outputting incorrect answers) is destructive to the construction of students' knowledge systems. To address this issue, this project adopts a Retrieval-Augmented Generation (RAG) architecture, using India's NCERT (National Council of Educational Research and Training) textbooks as authoritative knowledge sources, and ensures the accuracy and reliability of answers through multi-layer filtering mechanisms.

The system workflow consists of four core stages:

1. Knowledge Ingestion and Document Parsing: Extract text from NCERT textbooks using the pdfplumber library;
2. Semantic Chunking and Vectorization: Split text into semantic paragraphs and generate normalized vector embeddings via SentenceTransformer;
3. FAISS Index Construction and Similarity Retrieval: Use Meta's open-source FAISS library to build an IndexFlatIP index, supporting fast retrieval of relevant paragraphs using cosine similarity;
4. Generation and Filtering: Feed retrieval results into GPT-4o-mini to generate answers, apply multi-layer filtering such as confidence scoring and keyword overlap verification, and reject answers with low confidence.

Compared to traditional RAG systems, this project has three key improvements in reducing hallucinations:

1. Strict Retrieval Filtering: Only highly relevant retrieved content enters the generation stage to avoid model speculation;
2. Confidence Gating Mechanism: Set strict thresholds, and answers below the threshold are automatically filtered out;
3. Keyword Overlap Verification: Check whether key concepts in the answer exist in the original textbook content to prevent fabricated information.

The project uses a combination of mature tools from the Python ecosystem:

• Vector Retrieval: FAISS provides efficient similarity search;
• Text Embedding: SentenceTransformers generates semantic vectors;
• Large Language Model: OpenAI GPT-4o-mini is responsible for answer generation;
• Document Processing: pdfplumber parses PDF textbooks;
• Numerical Computing: NumPy and PyTorch support vector operations. The technology selection follows the principle of pragmatism, choosing tools that are proven mature and have active communities.

Suitable scenarios for the system:

• After-class Q&A: Explain content based on authoritative textbooks;
• Concept Explanation: Generate easy-to-understand explanations combined with textbooks;
• Homework Assistance: Quickly query knowledge points;
• Self-directed Learning: Support exploring textbooks at one's own pace. The built-in no-answer fallback mechanism cultivates healthy AI usage habits. When the AI says "I don't know", students need to consult materials or ask teachers, avoiding blind acceptance of incorrect answers.

Current limitations: Only supports NCERT textbooks. Future directions:

• Expand to more textbook systems and subject areas;
• Introduce multimodal capabilities to support non-text content such as charts and formulas;
• Add personalized learning path recommendations;
• Develop teacher-side tools to support custom knowledge base uploads.

The curriculum-grounded-rag-qa project demonstrates the pragmatic application of RAG technology in education, focusing on solving AI reliability issues. Through strict retrieval filtering, confidence gating, and keyword verification, it provides a reference implementation paradigm for the reliability of educational AI.