Vidhya-Setu

An offline AI tutor that teaches the prerequisites first, not just the answer.

GraphRAG based adaptive tutoring for NCERT Science, built from two Class 9 chapters. Runs on one ordinary laptop. No internet, no cloud API, no running cost.

---

The idea

A student who cannot follow acceleration usually has a gap further back, in velocity or in speed. Answer the question they asked and the gap stays. Next week it breaks something else.

So Vidhya-Setu does not start at the answer. Ask it about acceleration and it builds this path first:

mass  ->  motion  ->  speed  ->  velocity  ->  momentum  ->  force  ->  net force  ->  acceleration

Nobody wrote that chain by hand. The system read the textbook, worked out which concept depends on which, and stored it as a knowledge graph. The chain is a walk through that graph.

Then it teaches one step at a time and watches the replies. Say "I don't understand" and it goes backwards to an easier concept and tries a simpler analogy. Say "got it" and it moves on.

---

Why not just use a chatbot

What matters in a school	Cloud chatbot	Vidhya-Setu
Answers the question asked	Yes	Yes
Finds the gap behind the question	No	Yes
Works with no internet	No	Yes
Runs on an old school desktop	No	Yes
Student data leaves the building	Yes	Never
Cost per student, per year	Subscription	Nothing
Answers stay tied to the textbook	No	Yes, every explanation is grounded in NCERT text

---

From a PDF to a tutor

This part runs once, on a developer machine. Point it at a chapter PDF and it does the rest by itself.

flowchart LR
    A["NCERT PDF"] --> B["Clean text<br/>301 passages"]
    B --> C["Model reads each passage<br/>and names the prerequisites"]
    C --> D["Knowledge graph<br/>175 concepts, 197 links"]
    B --> E["Search index<br/>301 vectors"]
    D --> F["The tutor"]
    E --> F

Loading

The hard step is the third one. A small model asked for free-form JSON returns broken JSON often enough to ruin a batch, so the extractor uses a grammar that makes invalid output impossible while the model is still generating. Anything the model invents that is not actually in the passage gets dropped afterwards, and circular links are removed before the graph is saved.

Three commands, about twenty minutes, and a new chapter is live.

---

How a lesson runs

This is what happens on the school machine, with no network.

flowchart TD
    Q["Student asks a question"] --> R["Hybrid search<br/>vectors + BM25 keywords"]
    R --> C["Match it to a concept<br/>in the knowledge graph"]
    C --> P["Build the prerequisite chain<br/>capped at 8 steps"]
    P --> E["Explain the current concept<br/>Phi-3 Mini, offline"]
    E --> S{"Read the reply"}
    S -->|understood| A["Move forward one concept"]
    S -->|confused| B["Step back one concept<br/>and use a simpler analogy"]
    S -->|unclear| D["Explain the same concept again"]
    A --> E
    B --> E
    D --> E

Loading

The session is saved after every exchange, so a power cut costs at most one turn, and the student picks up where they left off.

---

Quick start

git clone https://github.com/Swaraj-Mandre/Vidhya-Setu.git
cd Vidhya-Setu
python -m venv .venv
.venv\Scripts\activate.bat      # Windows
source .venv/bin/activate       # macOS or Linux
pip install -r requirements.txt

Download the model, 2.4 GB, one time only:

python -c "from huggingface_hub import hf_hub_download; hf_hub_download(repo_id='microsoft/Phi-3-mini-4k-instruct-gguf', filename='Phi-3-mini-4k-instruct-q4.gguf', local_dir='models')"

Copy .env.example to .env, build the search index, start the tutor:

python scripts/build_vectorstore.py
python src/ui/app.py

Open http://127.0.0.1:7860. Every browser tab is a separate student.

---

Commands

Command	What it does	Time
python src/ui/app.py	Start the tutor	10 s to load
python scripts/run_checks.py	Run the full test suite, no model needed	under 1 min
python scripts/build_vectorstore.py	Rebuild the search index	30 s
python scripts/run_ingestion.py	Turn PDFs into text chunks	1 min
python scripts/run_kg.py	Rebuild the knowledge graph	17 min on CPU
python scripts/run_agent_test.py	End to end smoke test	2 min

To add a chapter: put the PDF in data/raw_pdfs/, add its chunk file to CHUNK_FILES in src/config.py, then run ingestion, graph and index in that order.

---

What is inside

Three things do the real work. Everything else is code that builds them or reads them.

Piece	What it is	Size
The knowledge graph data/graph/kg.json	175 concepts joined by 197 "learn this one first" links, pulled out of the textbook by the model	67 KB
The search index data/vectorstore/	301 passages of NCERT text, stored so the system can find the right one in milliseconds	0.63 MB
The language model models/*.gguf	Phi-3 Mini, writes the actual explanations. Downloaded separately	2.4 GB

Currently built from Class 9 Science, Chapter 4 (Describing Motion) and Chapter 6 (How Forces Affect Motion).

Is the graph any good? It was written by a model, so that is a fair question. Three things say yes:

• No circular logic. The graph is a proper DAG, so a concept can never end up being its own prerequisite.
• 43 links are backed by two or more separate passages in the textbook, not a single lucky sentence.
• The busiest concept is force, with 24 links. For two chapters about motion and forces, that is exactly what it should be.

The strongest links it found, ranked by how many passages support each one:

Learn this first	Before this	Passages agreeing
velocity	acceleration	28
force	acceleration	16
mass	force	14
motion	force	9
velocity	velocity-time graph	7

Every link is listed in kg_audit.json, with a smaller hand-checkable sample in kg_sample_review.json.

---

Stack

Layer	Choice	Why this one
Language model	Phi-3 Mini 3.8B Q4_K_M via llama-cpp-python	Small enough for CPU, good enough to teach
Structured output	GBNF grammar	Broken JSON becomes impossible during sampling, not just unlikely
Knowledge graph	NetworkX	Pure Python, and the whole graph saves to a 67 KB JSON file you can read by eye
Search	numpy + rank-bm25	301 passages is small enough for exact search. Two plain files, no database to migrate later.
Embeddings	BGE-small-en-v1.5	384 dimensions, strong on short questions
Ranking	Reciprocal rank fusion	Merges keyword and meaning results by rank, because their scores are not comparable
PDF parsing	pymupdf4llm	Keeps section headings, which the extractor uses as hints
Interface	Gradio	Browser UI with no frontend build step

Nothing reaches the network at runtime. The full test suite passes with HF_HUB_OFFLINE=1 and TRANSFORMERS_OFFLINE=1 set.

---

Hardware

Measured on Windows 11, CPU only, 4 threads.

Resource	Needed
RAM, idle	3.9 GB
RAM, peak while explaining	5.1 GB
Recommended	8 GB RAM
Disk	2.4 GB, nearly all of it the model
GPU	Not required
Internet	Not required
Speed	About 15 s to write a 120 word explanation

---

Roadmap

• Class 9 Science, Chapter 4 and Chapter 6
• Automatic knowledge graph from raw PDFs
• Backtracking tutor loop
• Multiple students at once
• Rest of the Class 9 syllabus
• Class 10 Science
• Hindi and Marathi
• Teacher view showing where a class is getting stuck
• One click installer for school machines

---

License

Copyright (C) 2026 Vidhya-Setu. Released under the GNU Affero General Public License v3.0.

In plain terms:

• Schools and students: use it, run it, change it, share it. Free, forever.
• Developers: fork it and build on it, as long as your version stays open under the same license.
• What you cannot do: take this work, close the source and sell it as your own product. That applies to running it as a hosted service too.

Every library used here is compatible with this license.

NCERT textbook content belongs to the National Council of Educational Research and Training, Government of India. Source PDFs are not included in this repository. Download them from ncert.nic.in.