🧠 VIKA – Vision Knowledge Assistant

VIKA is a lightweight, local-first Retrieval-Augmented Generation (RAG) system that transforms your PDF documents into a searchable knowledge base and provides accurate, source-cited answers via a local LLM.

⚡️ No cloud dependencies – everything runs locally on CPU or GPU
🧾 Ideal for lecture notes, research papers, or any scientific documents

---

🚀 Features

• 📂 Upload PDF documents via a web UI
• 🔍 Perform semantic search using FAISS
• 🧠 Rerank results with a local cross-encoder
• 🤖 Query a quantized Mistral-7B-Instruct LLM
• 🧾 Source citations shown in answers
• 🔁 Incremental indexing (no reprocessing needed)
• 🧑‍🏫 Optimized for student tutoring and educational QA

---

📦 Project Structure

.
├── app.py              # Gradio interface with sidebar, chat, and upload
├── chunker.py          # Text splitter (LangChain text splitter)
├── document_intake.py  # Deduplicates and extracts text from PDFs
├── embed_faiss.py      # Embeds chunks into FAISS index
├── retriever.py        # Vector search over FAISS index
├── reranker.py         # Reranks search results with a cross-encoder
├── prompt_builder.py   # Builds the final prompt with citations + chat history
├── generate.py         # Runs the LLM locally and saves response
├── pipeline.py         # Orchestrates the pipeline manually
├── parser_utils.py     # OCR/text extraction utilities (PaddleOCR, PyMuPDF) 
└── data/
    └── pdfs/           # Canonical store of processed PDF files and metadata

---

🛠️ Requirements

pip install -r requirements.txt

Key Dependencies

• sentence-transformers
• faiss-cpu
• torch
• langchain
• llama-cpp-python
• gradio
• pdf2image, fitz, paddleocr
• huggingface_hub

---

🖥️ Running the App

python app.py

Then open http://localhost:7860 in your browser.

---

🧾 How It Works

1. Upload PDFs
   Deduplicated by SHA-256, then text is extracted using PyMuPDF or PaddleOCR fallback.

2. Chunking
   The text is split into overlapping segments using a recursive splitter.

3. Embedding
   Chunks are embedded using all-MiniLM-L6-v2 and added to a FAISS index.

4. Retrieval + Reranking
   The system retrieves the top-k chunks via FAISS and reranks them using cross-encoder/ms-marco-MiniLM-L-6-v2.

5. Prompt Building
   Results are formatted using a Jinja2-based prompt builder, with human-readable source citations.

6. Generation
   A quantized Mistral-7B-Instruct model is used to generate the final answer locally.

---

🧪 Testing Manually

Full Index Build

python embed_faiss.py ./data/pdfs --out ./data/index --full

Retrieval + Reranking + Prompt

python retriever.py --query "What is a CNN?" --k 20 > hits.json
python reranker.py --query "What is a CNN?" --input hits.json --top 5 > top_hits.json
python prompt_builder.py --query "What is a CNN?" --hits top_hits.json > prompt.txt

Generation (CLI)

python generate.py

---

📌 Notes

• Files are stored by SHA-256 to avoid duplication.
• Original filenames are preserved in manifest.csv and shown in the UI.
• Supports both CPU and GPU execution.
• Gracefully handles empty indexes or missing model files.

---

📚 Example Use Case

Upload your course PDFs, ask "What is gradient descent?" and receive a cited, accurate answer directly grounded in your own material.

---

👨‍🔧 Future Improvements

• Add support for extracting/captioning images from PDFs
• Extend language support beyond English
• Add notebook support (.ipynb → markdown chunks)

---

🧠 Credits

Created with ❤️ for local-first, citation-focused document QA.
System name: VIKA – Vision Knowledge Assistant