📄 Document - RAG

A production-style Retrieval-Augmented Generation (RAG) application for querying enterprise PDF documents with grounded, source-cited answers.

Stack: FastAPI · Streamlit · LangChain · FAISS · SQLite · HuggingFace all-MiniLM-L6-v2 · Groq LLaMA-3.

---

✨ Features

• Upload one or many PDFs and ingest them into a persistent vector index.
• Strict dual-database architecture: FAISS for similarity, SQLite for metadata, chunks, and chat history.
• Semantic retrieval → metadata join → augmented prompt → LLM generation.
• Every answer shows source citations: filename, page, similarity, preview.
• Per-document delete, full knowledge-base reset, top-K tuning, multi-document QA.
• Modular service layout — clean separation between API, services, storage.

---

🏗 Architecture

              ┌──────────────┐        ┌──────────────────┐
PDF upload ──▶│ pdf_service  │──────▶ │ embedding_svc    │
              └──────────────┘        └────────┬─────────┘
                                               │ vectors
                                               ▼
                              ┌────────────────────────────┐
                              │ FAISS (IndexIDMap+FlatL2)  │  ◀── vector_id
                              └────────────────────────────┘
                                               ▲
                              ┌────────────────┴───────────┐
                              │ SQLite: documents, chunks, │
                              │         chat_history       │
                              └────────────────────────────┘
                                               ▲
                                               │ chunk_text + filename + page
User query ─▶ embed ─▶ FAISS search ─▶ SQLite lookup ─▶ prompt ─▶ Groq LLaMA-3 ─▶ answer + sources

---

📁 Project Structure

rag-docqa-pro/
├── backend/
│   ├── api/                 FastAPI app + Pydantic schemas
│   │   ├── main.py
│   │   └── schemas.py
│   ├── services/            Business logic
│   │   ├── pdf_service.py
│   │   ├── embedding_service.py
│   │   ├── ingestion_service.py
│   │   └── retrieval_service.py
│   ├── database/
│   │   └── sqlite_db.py     SQLite tables + queries
│   ├── vectorstore/
│   │   └── faiss_store.py   FAISS wrapper (persistent)
│   └── utils/
│       ├── config.py
│       └── logger.py
├── frontend/
│   └── streamlit_app.py     Streamlit UI (talks to FastAPI)
├── database/                rag_metadata.db (auto-created)
├── vectorstore/             faiss_index/    (auto-created)
├── requirements.txt
├── .env.example
└── README.md

---

🚀 Setup

# 1. Clone & enter
cd rag-docqa-pro

# 2. Create venv
python -m venv .venv
source .venv/bin/activate           # Windows: .venv\Scripts\activate

# 3. Install
pip install -r requirements.txt

# 4. Configure env
cp .env.example .env
# Edit .env and set GROQ_API_KEY (get one free at https://console.groq.com)

# 5. Run backend (terminal 1)
uvicorn backend.api.main:app --reload --port 8000
#   Swagger UI: http://localhost:8000/docs

# 6. Run frontend (terminal 2)
streamlit run frontend/streamlit_app.py
#   Open: http://localhost:8501

---

🧠 Workflow

1. Upload PDF in the sidebar → backend extracts pages with pdfplumber, splits into 500-char overlapping chunks, embeds with MiniLM, and stores:
   • the vectors in FAISS (keyed by vector_id)
   • the chunk text + filename + page in SQLite.
2. Ask a question → query is embedded → top-K similar vector_ids returned by FAISS → SQLite resolves them to the original text + source.
3. The retrieved context is injected into a system prompt and sent to Groq LLaMA-3 via LangChain.
4. The answer + structured sources (filename, page, similarity, preview) are returned and rendered, and the exchange is persisted to chat_history.

---

🔌 REST API

Method	Path	Purpose
GET	/health	Backend + index status
POST	/upload	Upload & ingest a PDF
POST	/query	RAG question → answer + sources
GET	/documents	List indexed documents
DELETE	/documents/{id}	Delete a document (+ vectors)
GET	/chat-history	Audit log (filter by session_id)
DELETE	/reset	Wipe knowledge base + history

Full interactive docs at http://localhost:8000/docs.

---

🖼 Screenshots

Add screenshots of the Streamlit UI here once running locally.

• docs/screenshot-chat.png
• docs/screenshot-sidebar.png

---

🛡 Security & Robustness

• Only .pdf uploads accepted (MIME + extension check).
• Temp files cleaned in a finally block.
• All SQLite writes guarded by a process-level lock.
• FAISS uses persistent IndexIDMap so vector IDs stay consistent across runs.
• Secrets read from environment variables only — never hard-coded.

---

🔭 Future Improvements

• Per-user authentication and session isolation.
• Hybrid retrieval (BM25 + dense) and re-ranking with a cross-encoder.
• Streaming token-by-token responses from Groq.
• Background ingestion queue for large PDFs.
• Postgres + pgvector for multi-tenant deployments.
• OCR fallback for scanned PDFs (Tesseract).
• Dockerfile + docker-compose for one-command deploys.