AI tutoring in Telugu and Hindi: how NCERT-aligned RAG actually works

Ask a generic AI chatbot to explain photosynthesis and it will happily oblige - in fluent, confident, textbook English that may or may not match what your Grade 8 CBSE student is actually studying this week. For a classroom in Bharat, that is two problems at once: the answer is in the wrong language, and it is grounded in the wrong curriculum.

Pravachak solves both with a technique called retrieval-augmented generation (RAG), tuned specifically for Bharat's boards and Bharat's languages. This post explains, without jargon, how it works - and why "grounded in NCERT" is a feature you should insist on before putting any AI in front of students.

The problem with a generic chatbot in a classroom in Bharat

A large language model on its own is a brilliant generalist and an unreliable specialist. It has read a lot of the internet, but it has not read your Class 8 Science textbook, it does not know your board sequences chapters in a particular order, and it does not know that your student thinks in Telugu even when they type in Roman script.

Left to its own devices, a generic model will:

• Answer from the wrong syllabus. It may explain a concept the way a US Common Core text does, using examples and terminology your student has never seen.
• Drift off the chapter. Ask about the water cycle and it may wander into thermodynamics the student will not meet for three more years.
• Default to English. Even when asked in Hindi, many models reply in English or in stilted, machine-translated Hindi that a child struggles to read.

None of this is malicious. It is simply what happens when a model has no grounding - no trusted source of truth to anchor its answer to.

What RAG does differently

RAG flips the process. Instead of letting the model answer from memory, Pravachak first retrieves the relevant passages from a trusted corpus, then asks the model to answer using only those passages. "Augmented" means the model's answer is augmented with real, cited source material before it ever reaches the student.

Here is the loop, step by step, for a Grade 8 CBSE student studying in Telugu who types "Photosynthesis ela jarugutundi?"

1. Normalise the input. The student wrote Telugu in Roman letters (code-mixing is how real students type). Pravachak transliterates it back to native Telugu script so the meaning is preserved for search.
2. Search the right shelf. Pravachak queries a curriculum index filtered to exactly this student's context - board: CBSE, grade: 8, subject: Science - so it can only pull from the correct chapter. This is a hybrid search: keyword matching and semantic (meaning-based) matching, so it finds the right passage even when the wording differs.
3. Ground the answer. The top matching passages from the NCERT chapter are handed to the model as source material, with an instruction: answer from this, cite the chapter, and reply in Telugu.
4. Reply in the student's language. The explanation streams back in native Telugu script - not translated after the fact, but generated in Telugu from the start - and names the chapter it came from.
5. Check for understanding. Rather than dumping the answer and moving on, the tutor pauses with a Socratic question - "In your own words, what role does chlorophyll play?" - before continuing.

The result is an answer that is on-syllabus, in the child's language, and traceable back to a real textbook chapter.

Why native-script, language-first matters

It is tempting to treat language as a translation layer bolted on at the end. Pravachak treats it as a first-class field on every AI request. Tutoring, quiz text, lesson plans and feedback are all generated in the learner's preferred language at native-script quality - Hindi, Telugu, Tamil, Kannada, Malayalam, Bengali, Marathi and English at launch, with more validated by native speakers before release.

This matters because comprehension collapses when a child has to translate in their head before they can even begin to reason. A student who understands the science but not the English sentence looks, on paper, like a student who does not understand the science. Language-first tutoring removes that false signal - and it is the difference our partner schools describe most often: students stop being afraid to ask questions, because now they can ask in the language they think in.

Grounding is also a safety feature

There is a governance dividend to RAG that school leaders care about. Because every answer is anchored to your curriculum corpus, the tutor is far less likely to invent facts ("hallucinate") or to wander into material that is off-syllabus or age-inappropriate. The model is constrained to teach from the shelf you approved.

Pravachak layers additional guardrails on top: sensitive personal information in a student's message is masked before the request leaves the school's environment, and every interaction is logged so administrators have a complete, auditable record of how AI is being used across the school.

What this means for your school

If you are evaluating AI tutors, the single most important question is not "how smart is the model?" - it is "what is the model grounded in?" A tutor grounded in your board's NCERT chapters, answering in your students' languages, is a fundamentally different product from a clever chatbot with a school logo on it.

That grounding is exactly what makes the outcomes measurable, too: because the tutor knows which chapter a student struggled with, teachers see real concept gaps, not vague engagement metrics.

Want to see NCERT-aligned RAG running on your own syllabus, in your own medium of instruction? Book a 30-minute demo and we will walk your team through a live session. Curious where it fits for a school your size? Our pricing is built around per-student seats across four tiers, with a free single-class pilot before you commit.