Understanding RAG: A Game-Changer for Language Learning

By Anne Pham and Hillary Zhong, Dickinson College

DOI: https://www.doi.org/10.69732/MQIK7847

Introduction: A STEM Student’s Perspective on AI in Education 

As a STEM student, I’ve been thinking a lot about AI and what it can do for education. In my math classes, AI is doing better and better – it’s gradually giving more accurate answers. Sometimes it messes something up, but you just need to come back and touch it up a bit, and it gives amazing results. A lot of classes are concept-based and theoretical, and AI can handle these well. 

However, my main concern is that students won’t learn anything with the presence of AI – we’ll just copy and paste and get around submitting homework quickly. But for language learning, it’s different. I’m positive that AI can excel when it comes to improving learning for students. It can act as an imaginary friend to do role play with students. And in language learning, teachers can assign homework with RAG, and students can engage in meaningful practice.

1. What is RAG and Why Does It Matter? 

RAG stands for Retrieval-Augmented Generation. It was first introduced by Lewis et al. (2020) as a way to combine retrieval systems with generative language models for knowledge-intensive tasks.Think of it as giving a chatbot a textbook, notes, or any documents you want it to reference before answering your questions. Instead of relying solely on its pre-trained knowledge (which might be outdated or too general), a RAG system retrieves relevant information from your specific documents and uses that to generate more accurate, personalized responses. 

Why Can RAG Be Useful in General? 

Our team built a RAG-based chatbot feature for language-learning courses at Dickinson College, which uses course textbooks and materials as reference documents so the chatbot’s responses stay grounded in what students are actually learning. Before our RAG feature was implemented, students often had trouble understanding the chatbot due to the difficulty of the generated text messages. Our RAG aims to fix this by providing context to the chatbot, so it can use language similar to that of the textbook. Additionally, if a professor wants to characterize the chatbot – such as making it a famous celebrity – without the use of RAG, the chatbot might hallucinate and give incorrect information. 

RAG offers several key advantages that make it particularly powerful for educational purposes, as outlined by Lewis et al. (2020) and subsequent research on retrieval-augmented systems (Gao et al., 2023): 

• Improved accuracy – the AI grounds its responses in your actual materials rather than making educated guesses based on its general training, giving you answers directly relevant to what you’re studying. 
• Customization control – you decide exactly what information the AI has access to, ensuring responses align with your curriculum or learning goals. 
• Up-to-date information – you can provide the AI with the latest materials, even if they weren’t included in its original training data. 
• Reduced hallucinations – by requiring the AI to reference specific documents, it’s much less likely to invent facts or provide misleading answers. 

RAG in Language Learning 

In language learning specifically, RAG transforms AI from a generic conversation partner into a knowledgeable tutor who has read your course materials. Imagine having a study partner who has memorized your entire textbook, all your class notes, your teacher’s handouts, and is available 24/7 to practice with you. That’s essentially what RAG provides. 

This means students can practice conversations, grammar, and vocabulary based on exactly what they’re learning in class. It means teachers can create customized learning experiences that align perfectly with their curriculum.

2. How RAG Is Used for Both Students and Educators

Promoting Self-Guided Learning in Students 

Conversation Practice 

Upload your vocabulary lists, grammar notes, or that dialogue you need to memorize, and practice with the AI. Unlike generic chatbots that might use words you haven’t learned yet, a RAG-powered system can roleplay scenarios using vocabulary and structures from your actual lessons. 

Personalized Explanations 

Struggling with a grammar concept? Upload your textbook chapter or class notes, and ask the AI to explain it differently. Since it has your specific materials, it can reference the exact examples your teacher used and provide alternative explanations that build on what you already know. 

Study Session Companion 

Before a test, upload all your study materials. The AI can quiz you, create practice exercises, and answer questions – all based on what will actually be on your exam, not generic language content. 

Creating Better Learning Experiences with RAG 

NotebookLM: Useful for Educators 

Google’s NotebookLM is an excellent example of RAG in action for language teachers. As highlighted by Meguro and Bryant (2025) in “AI Chatbots for Beginning Language Learners” , NotebookLM allows you to upload course materials and then interact with them in sophisticated ways. 

Teachers can use NotebookLM to: 

• Generate lesson plans: Upload your curriculum standards and textbook materials, and NotebookLM can help create structured lesson plans aligned with your content. 
• Create customized exams: RAG tools can help generate quiz questions and assessments based on your materials. A recommended approach is to ask the AI for more questions than needed – organized by section – and then manually select and modify the output to match the appropriate length and difficulty for your class. Since AI understands markup language, you can structure your prompt to request specific sections. For example: 

###Listening Comprehension### 

Generate 15 questions based on this audio transcript using vocabulary appropriate for first-year Chinese. Try to limit vocabulary to the chapters provided. 

##Grammar### 

Generate 10 fill-in-the-blank exercises based on the grammar patterns in Chapter 3. 

This approach ensures the output is already organized by section, making it easy to select and adapt questions for your exam. This keeps the teacher in control while dramatically speeding up the drafting process. 

• Develop differentiated materials: Create multiple versions of exercises at different difficulty levels, all drawn from the same source material. 
• Prepare discussion guides: Generate thought-provoking questions and conversation starters based on reading passages or cultural content you’ve shared. 

One particularly innovative feature is NotebookLM’s ability to create an “Audio Overview” – it can generate a podcast-style discussion between two AI hosts about your uploaded materials. This can be incredibly useful for providing listening practice in the target language or for creating engaging review materials. 

Suggestions on Addressing Limitations and Improving Performance 

NotebookLM is very good at referencing pages every time it answers, which helps ensure accuracy and allows users to verify information. However, it can take a long time to process and respond, especially when you upload large documents. In testing with a full ~500-page textbook, wait times were significant. 

An important limitation to be aware of is that NotebookLM does not reliably understand page numbers or chapter references within a fully uploaded textbook. For example, if you ask it to review vocabulary based on Chapter 7, it may not handle that request well when given the entire book at once. The workaround is to upload only the chapter or section you need – if you upload just Chapter 7, it will handle chapter-specific prompts much more effectively. For best results, upload smaller, focused documents such as individual chapters rather than entire textbooks. 

A second significant drawback is NotebookLM’s difficulty processing non-Latin scripts. For teachers working with Chinese, Arabic, Japanese, Korean, or other non-Latin writing systems, NotebookLM often struggles to accurately retrieve and reference content written in those scripts. This is a significant gap for language courses where the primary teaching material is written in the target language itself. For courses taught primarily in a non-Latin script, educators may want to explore alternative RAG tools – such as the chatbot system described in Section 3 – which handle non-Latin characters more reliably. 

Assignment Design with RAG 

Professors can assign homework that requires students to upload specific readings and then complete tasks using a RAG system. For example: “Upload Chapter 5 and have a 10-minute conversation about weekend plans using only the past tense structures we learned.” This ensures students are practicing the right content while preventing them from simply copying generic AI responses.

3. How RAG Improves Chatbots 

The Problem with Standard Chatbots 

When a student uses a regular chatbot like Gemini or ChatGPT without uploading documents, it responds based on its general training and the chatbot’s prompt. The student might encounter several issues: the AI might use vocabulary and grammar structures the student has not yet learned, reference cultural content that does not match the study material, or direct the conversation in a direction that does not fit the course curriculum. 

Our RAG System: A Separate Alternative 

Beyond tools like NotebookLM, our team has developed its own RAG-powered chatbot system tailored specifically for classroom language learning. Importantly, our code is free and downloadable, making it accessible to any institution willing to set it up. The two tools serve different purposes and have different strengths, as summarized below:

Feature	Our RAG Chatbot	NotebookLM
Non-Latin script support	Designed for it	Limited / unreliable
Curriculum alignment	Upload textbook chapters	Upload course materials
Page citation	Not currently supported	Cites page numbers
Character roleplay	Supported	Limited
Response speed	Fast	Slow for large docs
Cost	Varies by plan	Free tier available

How Our RAG Works 

When a professor uploads a file to our chatbot, the process looks like this: 

• The uploaded document is processed and broken into searchable chunks. 
• When the student asks a question, the system searches through these chunks for relevant information. 
• The most relevant sections are retrieved. 
• The AI generates a response that incorporates the retrieved information with its general language understanding. 
• The result is grounded in the uploaded documents rather than being completely generated from scratch. 

The RAG Difference: A Practical Experiment 

Consider a practical example using our chatbot with a Chinese language textbook as the RAG source. The character Anna is a friendly Chinese student helping a foreign exchange student practice Mandarin. 

Without RAG (prompt only): 

The chatbot generates conversations using whatever vocabulary and grammar it deems appropriate – often advanced words, complex sentence patterns, or cultural references beyond the student’s current level. 

Student: 你好。北京真美啊。 (Hello. Beijing is really beautiful!) 

Anna (without RAG): 北京确实很美，有很多名胜古迹，比如故宫和长城。你去过北京的哪些地方 呢? (Beijing is indeed beautiful, with many landmarks like the Forbidden City and the Great Wall. Which places in Beijing have you visited?) 

With RAG (textbook uploaded): 

Now, with the textbook attached and instructions to use simple vocabulary similar to the examples in the uploaded materials, the difference is clear: 

Student: 你好。北京真美啊。 (Hello. Beijing is really beautiful!) 

Anna (with RAG): 是啊，北京很漂亮。你去过哪儿呢？ (Yes, Beijing is beautiful. Where have you been?) 

Notice how Anna stays within the vocabulary and grammar patterns from the textbook, creating an authentic conversation perfectly calibrated to the student’s current level. 

Strengths and Limitations of Our Current RAG System 

Our current RAG prioritizes either language learning or chatbot communication style, but not both simultaneously. For example, if we want the chatbot to resemble a celebrity or fictional character, it fetches files that indicate their personality. If we want the chatbot to speak in simple language similar to a textbook, the RAG fetches chunks with examples to mimic. One limitation is that the system cannot do both effectively at once, so professors must choose a priority. Additionally, due to the way documents are processed, the system cannot currently refer to a specific page number within a document, or identify the name of a specific document when the professor has uploaded more than one. 

The chatbot we developed cost $1.20 over the last month, with around 2 million tokens sent to OpenAI. It was able to use RAG at negligible cost (around 1 cent for an upload of a 1 MB file), while supporting students across 5 courses. Teachers utilized the full functionalities of the chatbots along with RAG functionalities at negligible cost compared to other programs with a similar purpose, such as Mizou. However, it is worth noting that running our chatbot requires a Linux webserver and some hosting expertise, which may be a barrier for educators without dedicated technical support.

4. Mizou: A Practical Test Case for AI Language Learning 

Overview of Mizou 

For educators who want to take advantage of AI-powered language tutoring without the technical overhead of hosting their own chatbot, Mizou offers a practical, ready-to-use alternative. To understand how RAG and AI-powered chatbots function in real-world language learning environments, I tested Mizou, an AI chatbot platform designed specifically for educational use. Mizou’s interface and features reveal both the practical strengths of modern AI language tutors and important considerations for educators. It is worth noting that while Mizou is more accessible from a setup standpoint, using its full RAG capabilities does come at a cost (discussed below), whereas our custom-built chatbot system is free and downloadable for those with access to a Linux webserver and hosting expertise. 

Key Features Observed 

Speech-to-Text Conversion 

One particularly useful feature is Mizou’s voice message functionality. When students send voice messages, the platform performs automatic speech-to-text conversion, displaying the transcribed text before sending it to the AI tutor. This serves multiple pedagogical purposes: it helps students see how their spoken language is being interpreted, provides immediate feedback on pronunciation clarity, and allows students to correct their transcription before the AI responds. 

Natural Corrections and Feedback 

During testing, Mizou demonstrated strong capability in providing corrections that feel natural and encouraging rather than harsh. The AI responds to student errors constructively, maintaining conversational flow while gently correcting mistakes. For example, when a student said “我很好，你呢?” (Wǒ hěn hǎo, nǐ ne?), which translates to “I’m doing well, and you?”, the chatbot responded naturally while modeling the correct pattern. 

Accent Quality 

The chatbot’s spoken responses feature natural-sounding Mandarin with appropriate tone and intonation. This is a significant advantage for students who benefit from hearing authentic pronunciation, as it reinforces proper tonal patterns that are critical in Chinese. 

How Teachers Create and Distribute Chatbots 

Mode 1: AI-Generated 

Teachers input their learning objectives, and Mizou automatically generates the chatbot prompt. This is ideal for teachers who want a quick, structured approach without manually writing detailed instructions. 

Mode 2: Custom 

Teachers write AI instructions from scratch, allowing for highly specific customization. Teachers can specify that the chatbot should roleplay as a specific character, restrict vocabulary to certain textbook chapters, emphasize particular grammar structures, or adopt a specific teaching methodology. Teachers have a maximum of 1,000 characters for detailed instructions. 

Mode 3: Hybrid (More Options) 

Beyond the basic modes, Mizou offers additional customization through “More options,” which allows teachers to blend generated and custom elements. 

The Knowledge File and RAG 

One of Mizou’s most powerful features is the Knowledge File upload section, which implements RAG. Teachers can upload textbook chapters, vocabulary lists, grammar notes, class handouts, dialogue scripts, and more. Currently, Mizou supports one file per chatbot. When a teacher uploads a knowledge file, the chatbot references that material, ensuring vocabulary, grammar structures, and content stay within the scope of what students have actually learned. 

Prompt Writing: Crafting Effective Instructions 

When using Custom mode, effective prompts typically include: character and personality definition, vocabulary and grammar constraints, teaching style and tone, conversation scope, response length expectations, and an error correction strategy. An example prompt: 

“You are Anna, a friendly Chinese student who recently became friends with me. We met about a month ago when I arrived for my study abroad program. Use ONLY vocabulary and grammar from pages 1–50 of our textbook. Speak only in Chinese-pretend you don’t know much English. Be warm, encouraging, and patient. When I make mistakes, correct them naturally in Chinese. Ask follow-up questions to keep our conversation flowing. Suggest activities like going to restaurants, the library, or shopping. Keep responses to 3–5 sentences unless I ask for more.” 

Cost Implications 

While Mizou offers excellent features, the Knowledge File upload functionality (which implements RAG) requires a paid plan. Full RAG capabilities fall under the premium Individual and Team plans ($14.90/month and $19.90/month respectively), making Knowledge File uploads inaccessible for free tier users. Teachers using free accounts can still create effective chatbots through custom prompts alone, but they cannot upload curriculum materials to ensure vocabulary alignment. 

Conclusion on Mizou 

Mizou demonstrates that modern AI chatbot platforms can deliver high-quality conversational practice with natural corrections and good accent quality. The voice transcription feature is particularly innovative for language learning. However, the paywall on RAG functionality and unclear support for full classroom management features suggest that while Mizou excels as a supplementary practice tool for individual learners, it may not yet be the comprehensive solution for teachers seeking to integrate AI tutoring into structured curricula without additional cost or complexity. 

Conclusion 

RAG represents a meaningful step forward in how AI can support language learning. By grounding chatbot responses in actual course materials, it transforms generic AI tools into personalized tutors that stay aligned with what students are studying and what teachers intend to teach. 

As this article has shown, there is no single solution that fits every classroom. NotebookLM offers strong citation support and easy access for educators comfortable uploading focused documents, but struggles with non-Latin scripts and large files. Our custom RAG chatbot handles non-Latin characters reliably and is free to download, but requires technical infrastructure to run. Mizou provides a polished, ready-to-use experience with innovative features like speech-to-text, but places its most powerful RAG capabilities behind a paywall. 

The right tool depends on a teacher’s priorities: technical resources, target language, budget, and how deeply they want to integrate AI into their curriculum. What is clear is that RAG-powered tools are already making AI more useful and more appropriate for language classrooms – and as these tools continue to improve, the gap between generic AI responses and genuinely curriculum-aligned practice will only narrow. 

For educators, the most important takeaway is that RAG puts you back in control. You decide what the AI knows, how it speaks, and what it teaches. Used thoughtfully, it has the potential to give every student access to a patient, knowledgeable, always-available practice partner – one that has actually read the textbook. 

For students, the experience of working with a RAG-powered chatbot is different from using a generic AI tool. Instead of an assistant that speaks with unpredictable vocabulary or wanders off-topic, students can lean on a resource that speaks their language within the exact scope of what they’ve learned. This lowers the intimidation factor: students aren’t afraid of encountering an unfamiliar grammar structure or being misled by an AI’s confident but incorrect answer. From our own experience testing these tools, this trust in the chatbot’s grounding makes it easier to engage in genuine, sustained practice rather than tentative, guarded interactions. For other students exploring RAG-based tools, our biggest takeaway is that these chatbots work best as a practice partner, not a shortcut—a resource to stress-test your understanding of vocabulary and grammar in a low-stakes way, rather than a replacement for actual study. Understanding why the chatbot responds the way it does (i.e., that it’s referencing your actual course materials) can also help students become more thoughtful, critical users of AI more broadly, both in language learning and beyond.

References

Gao, Y., Xiong, Y., Gao, X., Jia, K., Pan, J., Bi, Y., Dai, Y., Sun, J., & Wang, H. (2023). Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997. 

Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W., Rocktäschel, T., Riedel, S., & Kiela, D. (2020). Retrieval-augmented generation for knowledge-intensive NLP tasks. Advances in Neural Information Processing Systems, 33, 9459-9474.

Meguro, A. and Bryant, T. (2025). AI Chatbots for Beginning Language Learners. The FLTMAG. https://www.doi.org/10.69732/CZNP1530

AI Disclosure: AI was used for for spelling and grammar suggestions.