📘 Knowledge Assistant — RAG-powered Notebook

A NotebookLM-inspired AI research assistant. Upload a document — from any field of life (business, legal, medical, academic, personal, technical, etc.) — and ask questions about it. Answers are generated only from the content of your uploaded files using Retrieval-Augmented Generation (RAG).

✨ Features

• Multi-format ingestion: PDF, Word (.docx), Excel (.xlsx / .xls), CSV, TXT, and images (.jpg / .png, via OCR).
• Semantic search: fastembed (ONNX, no PyTorch) embeddings + FAISS vector index.
• Grounded answers: Groq (openai/gpt-oss-20b) generates answers strictly from retrieved context, with source attribution.
• Modern, modular UI: three-column NotebookLM-style layout (Sources & History / Chat / Studio) built with Streamlit + custom CSS.
• Clean, scalable codebase: parsing, chunking, embeddings, vector store, LLM calls, and UI are all separated into their own modules.

🧠 How it works (RAG pipeline)

1. Upload a file → file_parsers.py extracts raw text (OCR for images).
2. Chunk the text into overlapping ~220-word pieces (text_chunker.py).
3. Embed each chunk with all-MiniLM-L6-v2 (embeddings.py).
4. Index the vectors in a FAISS IndexFlatIP store (vector_store.py).
5. Ask a question → the question is embedded and the top-K most similar chunks are retrieved.
6. Generate → retrieved chunks are passed as context to Groq's openai/gpt-oss-20b model, which is instructed to answer only from that context.

🚀 Run locally

1. Clone and set up a virtual environment

git clone https://github.com/<Kamran-31>/knowledge-assistant.git
cd knowledge-assistant
python -m venv venv
source venv/bin/activate      # Windows: venv\Scripts\activate

2. Install system dependency for OCR (Tesseract)

3. Install Python dependencies

pip install -r requirements.txt

4. Add your Groq API key

Get a free key at console.groq.com, then:

cp .streamlit/secrets.toml.example .streamlit/secrets.toml

Edit .streamlit/secrets.toml:

GROQ_API_KEY = "gsk_your_real_key_here"

(Alternatively, paste the key into the ⚙️ Settings popover inside the running app — useful for quick testing without touching secrets.)

5. Run the app

streamlit run app.py

Open the local URL Streamlit prints (usually http://localhost:8501).

🛡️ Notes & limitations

• Answers are grounded strictly in uploaded content — if the document doesn't contain the answer, the assistant says so instead of guessing.
• Image OCR quality depends on image clarity; scanned/handwritten text may extract poorly.
• The in-app "Notebook History" and "Studio" tools reflect the current browser session only (no database is wired up) — everything resets on a full app restart. Swap in a persistent store (SQLite/Postgres/S3) if you need durability across sessions.
• Vector index is in-memory per session; very large documents will use more RAM — trim CHUNK_SIZE_WORDS/file size if you hit Streamlit Cloud's free-tier memory limits.