🎓 ABV-IIITM College RAG Chatbot

Ask natural-language questions about ABV-IIITM Gwalior — programs, fees, curricula, ordinances, hostel, scholarships — and get answers cited from the institute's own PDFs.

---

A complete Retrieval-Augmented Generation pipeline — crawl → extract → chunk → embed → index → retrieve → answer — that runs offline on CPU with no torch, no paid services, and no API keys for the core pipeline. Embeddings and reranking run on ONNX. Only the final answer step calls an LLM, and that can be a free hosted model (Groq) or a local one (Ollama).

💬 Example

Q:  What is the hostel mess fee for M.Tech students?
A:  The hostel mess fee for M.Tech students is ₹20,000.
    └─ source: Fees-Details-for-the-Students-2025-batch-…-2026.pdf

✨ Features

• 📊 Table-aware ingestion — fee structures and credit tables survive as real Markdown tables; scanned PDFs are OCR'd automatically.
• 🏷️ Metadata-filtered retrieval — every chunk is tagged with audience (BTech · MTech · MBA · MS · PhD · IPG) and topic (Fees · Curriculum · Hostel · …), so a query can be scoped to exactly "M.Tech fees".
• 🔀 Hybrid search — semantic vectors (FAISS) + BM25 keywords, fused with Reciprocal Rank Fusion → finds both paraphrased questions and exact identifiers like Rule 30.7.
• 🎯 Cross-encoder reranking (ONNX, opt-in) for top-result accuracy.
• 🔎 Grounded, cited answers — every answer traces back to its source documents.
• 🖥️ Clean Streamlit chat UI with program/topic filters.

🏗️ How it works

Eight standalone stages — each stage's output feeds the next:

1. download_pdfs.py        →  pdfs/           crawl the site + download PDFs
2. extract_pdfs_local.py   →  extracted/      PDF → Markdown (pymupdf4llm + RapidOCR)
3. chunk_documents.py      →  chunks.jsonl    recursive chunking + metadata tags
4. embed_chunks.py         →  embeddings.npy  embeddings (fastembed · bge-small-en-v1.5)
5. build_faiss.py          →  faiss_store/    FAISS index + docstore + config
6. hybrid_search.py        →  HybridRetriever (vector + BM25 + RRF + rerank)
7. answer.py               →  retrieve → grounded prompt → LLM → cited answer
8. app.py                  →  Streamlit chat UI

🧰 Tech stack (all free)

Stage	Library
Crawl	requests · beautifulsoup4
Extract / OCR	pymupdf4llm · rapidocr-onnxruntime
Chunk	langchain-text-splitters
Embed / rerank	fastembed (ONNX — no torch)
Vector index	faiss-cpu
Keyword search	rank-bm25
Answer LLM	Google Gemini (free tier) · Groq (free) · Ollama (local)
UI	streamlit

⚡ Quickstart

git clone https://github.com/itz-puneet/iiitm-rag-chatbot.git
cd iiitm-rag-chatbot

python -m venv .venv
.venv\Scripts\activate            # Windows   (macOS/Linux: source .venv/bin/activate)
pip install -r requirements.txt

cp .env.example .env              # then paste your free Gemini or Groq key into .env

Build the knowledge base (once, in order), then launch the app:

python download_pdfs.py --depth 2 --out ./pdfs
python extract_pdfs_local.py
python chunk_documents.py
python embed_chunks.py
python build_faiss.py

streamlit run app.py

💡 The first embed/search run downloads the ONNX models (~130 MB embedder, ~90 MB reranker) once, then works fully offline. On Windows, prefix console runs with PYTHONIOENCODING=utf-8 — the documents contain ₹ and Devanagari.

🔍 Usage (CLI)

# hybrid retrieval (add --rerank for the cross-encoder pass)
python hybrid_search.py --query "rule 30.7" --rerank

# a full cited answer, scoped with metadata filters
python answer.py --q "What is the hostel mess fee for M.Tech students?" --audience MTech --topic Fees

⚙️ Configuration (.env)

# Option 1 — Google Gemini API (Recommended, 100% free at https://aistudio.google.com/)
GEMINI_API_KEY=your_gemini_api_key_here
GEMINI_MODEL=gemini-2.5-flash

# Option 2 — Free hosted LLM (Groq, OpenAI-compatible at https://console.groq.com/)
# OPENAI_API_KEY=gsk_your_key
# OPENAI_BASE_URL=https://api.groq.com/openai/v1
# OPENAI_MODEL=llama-3.3-70b-versatile

Or leave those unset and run a local model with Ollama: ollama pull llama3.2.

📂 Project structure

download_pdfs.py         # 1  crawl + download
extract_pdfs_local.py    # 2  PDF → Markdown (+ OCR)
chunk_documents.py       # 3  chunk + tag metadata
embed_chunks.py          # 4  embeddings
build_faiss.py           # 5  FAISS vector store
hybrid_search.py         # 6  hybrid retrieval + rerank
answer.py                # 7  RAG answer layer
app.py                   # 8  Streamlit UI
requirements.txt · .env.example

📝 Notes

• Design constraint: the whole stack stays free and local, deliberately avoiding torch (ONNX via fastembed). Only the final answer step optionally calls a hosted LLM (Groq's free tier).
• Data is not committed — pdfs/, extracted/, chunks.jsonl, embeddings.npy, and faiss_store/ are large and reproducible; rebuild them with the pipeline above.
• Retrieval semantics: audience is a list (use a contains filter), and "General" means institute-wide — a program query matches audience ∈ {<program>, "General"}.
• Built for ABV-IIITM Gwalior, but adaptable to any institution by re-pointing the crawler and re-running the pipeline.

_{Built with a fully free, local RAG stack · embeddings & reranking on ONNX}