Gurukul AI brings NCERT-aligned tutoring to India

Gurukul AI brings NCERT-aligned tutoring to India

Introduction

If you’ve ever tried using a general AI chatbot for school questions in India, you’ve probably seen the problem: it’s smart, but it doesn’t feel like your textbook. That’s the core idea behind Gurukul AI, an interactive AI-driven educational platform designed specifically for the Indian education system—built from new research published here: the original paper.

The research (by Isha Narang, Sneh Gosai, and Mayank Singh) points out a mismatch that’s been bothering educators and students: big LLMs like ChatGPT and LLaMA are largely trained on Western-centric data, while Indian schooling is linguistically diverse, exam-structured, and tightly linked to standardized syllabi—especially NCERT for Classes 9–12. So the authors didn’t just “try” a chatbot. They built a localized system that actually speaks the language of Indian curricula—sometimes literally.

In short, the team curated a syllabus-aligned question-answer dataset directly from NCERT textbooks, then fine-tuned an LLM and wrapped it in a Retrieval-Augmented Generation (RAG) setup. The result is GurukulAI111—a platform where students can chat in English and Hindi, get doubts cleared, practice exam-style questions (MCQs + theory), and receive answers grounded in the relevant chapter. The dataset is publicly available at Hugging Face, and the hosted website code is on GitHub at lingo-iitgn/GurukulAI.

Why This Matters

Here’s what makes this research feel urgent and practically useful: Indian students don’t just need answers—they need answers that match the teaching style, sequence of concepts, and exam expectations of their syllabus. General-purpose AI can be accurate in a broad sense, but it often misses the “NCERT lens,” which is exactly what students rely on for marks.

A specific scenario where this can be used today: imagine a Class 10 student preparing for CBSE-style exams. They ask, “Why does carbon form four bonds?” A generic chatbot might explain bonding correctly—but it may frame it using a different explanation structure or bring in extra details that don’t align with the textbook. With GurukulAI, retrieval pulls the most relevant chapter passages from NCERT, so the answer is more likely to match what teachers and examiners expect. That reduces the annoying back-and-forth where students have to “translate” AI explanations into textbook language.

This work also builds on earlier AI-in-education efforts in a smarter way. Prior LLM education systems showed that conversational tutoring and feedback are promising, but many weren’t curriculum-aligned or didn’t support bilingual interaction in a grounded way. GurukulAI isn’t trying to replace tutoring or teachers—it’s trying to make AI auditable in context, using chapter-level alignment as the anchor. You can think of it like moving from “generic encyclopedia answers” to “a tutor who studied your exact textbook first.”

A Textbook-Aligned Dataset: The Real Starting Point

A lot of AI education projects start with the model. This paper starts with the content.

The Gurukul dataset: 18,720 NCERT-based Q&A pairs

The authors curated 18,720 question–answer pairs for Classes 9–12 across five subjects:
- Science
- Mathematics
- Social Science
- English
- Hindi

This is not just a dataset of random school questions. The key innovation is that it’s aligned to NCERT textbooks and includes chapter-level annotations. That matters because India’s school learning isn’t only concept-based—it’s also chapter-based. Students ask questions in the middle of a chapter, revise chapter summaries, and expect solutions to follow the textbook’s progression.

Where the content came from (and what they filtered out)

The team collected information from:
1. Official NCERT PDFs in English and Hindi
2. Previous year CBSE question papers (from public repositories)
3. Educational websites like Vedantu, Byju’s, LearnCBSE, and Tiwari Academy

They then parsed NCERT using layout-preserving extraction (PyMuPDF) and scraped dynamic websites with a blend of manual curation and automation (including Selenium). After extraction, they used Claude Sonnet API to:
- remove incomplete / diagram-dependent Q&As,
- rewrite ambiguous or compound questions into standalone ones,
- align phrasing with the NCERT teaching tone,
- fill missing parts and keep outputs coherent.

Finally, they manually reviewed entries for fluency, syllabus relevance, and coherence. One important consequence: they excluded visual-context Q&As (diagrams/charts) so the dataset stays text-grounded.

Why the “chapter tag” feature is a big deal

Even if two answers are factually correct, students care about whether the explanation matches their current chapter. Chapter tagging enables:
- retrieval to pull the correct relevant section,
- “source linking” back to an NCERT chapter,
- better practice modes (MCQ from a selected chapter),
- a more predictable learning experience than open-domain chat.

Fine-Tuning LLaMA + RAG: How GurukulAI Stays Grounded

Now for the brain part: how the system actually generates answers without drifting into generic chatbot territory.

The model choice and training approach

The authors fine-tuned LLaMA 3.1 8B using LoRA (Low-Rank Adaptation) with instruction-style prompts based on their curated dataset. Training ran on an NVIDIA A100 40GB GPU, with a 4,096-token context window.

Why LoRA? It lets you adapt the model efficiently without retraining the entire network from scratch—useful when your goal is curriculum style and syllabus alignment rather than totally new language capability.

Retrieval-Augmented Generation (RAG): a “textbook memory” layer

Fine-tuning teaches the model how to respond in a textbook-like manner. But RAG ensures the model answers based on what it fetched from the NCERT corpus.

Here’s the analogy the paper implicitly supports:
- Fine-tuning is like teaching a tutor your preferred explanation style.
- RAG is like handing the tutor a stack of your exact chapters and saying: “Use these pages when you answer.”

How retrieval works in practice

• NCERT chapters are chunked into coherent passages (about ~4,000 characters)
• Each chunk is embedded using BGE-m3
• Chunks are indexed in FAISS
• At inference time:
  1. user query is embedded,
  2. top 5 relevant chunks are retrieved,
  3. retrieved text + query are combined into a context-rich prompt,
  4. the fine-tuned LLM generates the final answer.

They also implemented KV caching and query deduplication to reduce latency, plus dynamic prompt templates to keep instruction-following consistent.

The four configurations comparison (does RAG actually help?)

They tested several model setups on the same test set to measure impact. Here’s how the system components stack up:

Their results showed that the full RAG model performs best, and removing retrieval caused measurable drops:
- Semantic similarity drops by 13%
- ROUGE-L drops by 6–8%

That’s strong evidence that retrieval-based grounding isn’t just a nice extra—it’s central to syllabus-aligned performance.

GurukulAI’s Interactive Features: More Than Q&A

The research isn’t only about model quality—it’s also about student interaction design.

Chat + bilingual support (English and Hindi)

Students can interact with the model via a web interface built with Flask. The frontend supports:
- chatting in English and Hindi,
- asking subject-specific questions,
- getting context-aware responses.

This bilingual support matters in India not as a “nice-to-have,” but because students often switch between language registers depending on comfort level.

Image-based doubt resolution (but note the scope)

One helpful feature: image-based question handling. Users can upload scanned or handwritten questions.
- For digital PDFs, text is extracted using PyMuPDF (Fitz).

Important limitation: the dataset excluded diagram-dependent Q&As during curation. So while the platform can accept images, the system’s learning corpus is primarily text-based—and the paper explicitly notes that multimodal visual inputs like diagrams/charts aren’t fully supported yet.

MCQ practice mode (chapter-selectable)

Students can select a chapter and switch into MCQ practice. The platform:
- generates MCQs using prompt-based inference over chapter summaries,
- explains and scores options in real time.

That’s a very exam-oriented workflow—exactly what Indian schooling demands.

Theory answer practice with feedback

For longer descriptive answers, the platform supports:
- generating descriptive questions from selected chapter content,
- evaluating free-form student responses,
- giving feedback and improvement suggestions.

This is where the platform becomes more like a study partner rather than just a Q&A bot.

Follow-up questions and learning continuity

After each interaction, the model suggests three contextually related questions. This encourages:
- revision,
- exploration of connected subtopics,
- less “single-shot tutoring” and more continuous learning.

Source linking: tying answers back to chapters

Another key educational feature is source linking—each answer is grounded in the most relevant NCERT chapter chosen via semantic retrieval. In real study habits, this helps students verify explanations and build trust.

What the Evaluation Found (and What Students Actually Thought)

The paper evaluates the system through both automated metrics and a real student study.

Automatic evaluation: 700 test questions across subjects

They used a held-out test set of 700 questions, covering:
- multiple subjects (Science, Social Science, English, Hindi),
- labeled difficulty levels: easy, medium, hard,
- ground-truth answers manually curated from NCERT textbooks.

They measured:
- ROUGE-L: token overlap with reference answers (useful for factual coverage)
- Semantic similarity: cosine similarity between sentence embeddings (captures contextual alignment)

Performance patterns they observed

• GurukulAI performs best on Social Science and Science, with high ROUGE-L and semantic similarity.
• English and Hindi are slightly lower, likely because language subjects allow more variation in acceptable phrasing.
• Performance stays relatively stable across Classes 9–12, with a small dip in Class 11, which they attribute to increased abstraction and complexity.
• Hard questions reduce scores: ROUGE-L drops more sharply, but semantic similarity remains relatively robust—suggesting answers are still contextually relevant even when not perfectly aligned in length or phrasing.

Human study: 43 students’ feedback

They ran a structured human study with 43 participants across Classes 9–12, plus undergraduates with past CBSE experience. Students interacted with the platform and then completed surveys on a 5-point Likert scale.

Key results (percentage agreement highlights):
- 86% found responses factually reliable
- 84% found it helpful
- 81% found it easy to understand
- 88% liked response time
- 89% said it improved their subject understanding
- 77% found it less distracting than conventional online search
- 72% preferred it over existing methods

Open-ended responses praised:
- UI,
- speed,
- textbook alignment,

and described GurukulAI as “a personal tutor aligned with our books.” Suggested improvements included simplifying some explanations and enhancing UI/UX consistency and follow-up features.

Comparing to general chatbots: why students preferred “syllabus grounding”

The paper notes that while some participants preferred broader LLMs like ChatGPT or Gemini for general queries, most favored GurukulAI specifically because it stays within syllabus scope and NCERT grounding.

That’s the heart of the research claim: localized tutoring isn’t just about language—it’s about curriculum constraints.

Limitations and the Next Steps GurukulAI Still Needs

No system is perfect, and the paper is pretty transparent about what’s missing.

Current coverage: only Classes 9–12 and NCERT-based

• The platform supports only Classes 9–12 (science stream is mentioned for coverage limitations).
• It is aligned only with NCERT, which limits applicability to ICSE and various state board syllabi.

Reasoning depth: multi-step problems are tough

They report struggles with high-cognitive-load questions, especially in Mathematics and Physics, due to the need for multi-step reasoning. (These subjects are also excluded from the portal right now, per the limitation notes.)

Multimodal gaps: diagrams and charts aren’t supported well

Despite image input for doubts, the system currently lacks strong capability for visual question types like diagrams and charts—consistent with the dataset being text-focused.

Hallucinations remain possible

Even with retrieval grounding, LLMs can still produce misleading content occasionally. The paper advises students to cross-verify responses—especially for critical exam submissions.

Ethics: fair-use dataset curation and no personal data collection

The dataset was curated from publicly available NCERT textbooks and educational websites under fair-use for non-commercial academic purposes. The project reports no personal user data collected during training or evaluation. They also mention potential bias persistence due to automated processing and that current language support is limited to English and Hindi.

Key Takeaways

• GurukulAI is built for Indian syllabi, not generic tutoring: it uses a NCERT-aligned dataset (18,720 Q&A pairs) with chapter-level tagging for Classes 9–12.
• The system combines fine-tuned LLaMA 3.1 8B with RAG using BGE-m3 embeddings and FAISS indexing, retrieving the top 5 relevant NCERT passages for each query.
• RAG matters: removing retrieval drops semantic similarity by 13% and ROUGE-L by 6–8%, showing that textbook grounding improves alignment.
• In evaluation on 700 test questions, performance is strongest in Social Science and Science, with more variability in language subjects due to acceptable phrasing differences.
• A human study with 43 students found high usefulness and trust: 86% reliability, 89% improved understanding, and 72% preferred GurukulAI over existing methods.
• The platform is promising but not finished: it currently lacks broad board coverage (ICSE/state), struggles with multi-step high-difficulty reasoning, and has limited support for diagram-heavy multimodal questions.

If you’re a student, teacher, or education builder, the big “aha” here is that curriculum alignment is not a tweak—it’s the product. GurukulAI is essentially proof that localized educational AI can feel more like a personal tutor when it’s anchored to the exact textbook logic students study from.