Form Processing RAG App

A Retrieval-Augmented Generation (RAG) pipeline built with Streamlit, PostgreSQL (with pgvector), and Llama 3 via Groq API. This app allows users to upload documents, extract and chunk text, embed with sentence-transformers, store in PostgreSQL, and retrieve relevant chunks for LLM-based Q&A.

---

Features

• Document Upload: Supports PDF, DOCX, images, and text files.
• Text Extraction: OCR for images, parsing for PDFs/DOCX/text.
• Chunking & Embedding: Splits text and generates 384-dim embeddings using sentence-transformers.
• Vector Database: Stores embeddings in PostgreSQL with pgvector extension.
• Semantic Search: Retrieves relevant chunks using vector similarity.
• LLM Q&A: Uses Llama 3 via Groq API for question answering over retrieved context.
• Named Entity Recognition: Extracts and displays entities in both JSON and human-readable formats.
• Data Management: Sidebar button to clear all data (for research phase).

---

Tech Stack

• Frontend: Streamlit
• Backend: Python
• Database: PostgreSQL (with pgvector, e.g., Supabase)
• Embeddings: sentence-transformers
• LLM: Llama 3 via Groq API
• OCR: Mistral API (for images)

---

Setup

1. Clone the repository:

   git clone https://github.com/Daramanohar/rag-docs-engine/tree/main
   cd rag-docs-engine

2. Install dependencies:

   pip install -r requirements.txt

3. Configure environment variables:

   • Add your API keys and database credentials to .streamlit/secrets.toml or the Streamlit Cloud secrets UI:

     MISTRAL_API_KEY = "your-mistral-key"
     GROQ_API_KEY = "your-groq-key"
     DB_HOST = "your-db-host"
     DB_NAME = "your-db-name"
     DB_USER = "your-db-user"
     DB_PASSWORD = "your-db-password"
     DB_PORT = "5432"

4. Enable pgvector extension in your PostgreSQL database (if not already enabled):

   CREATE EXTENSION IF NOT EXISTS vector;

5. Run the app:

   streamlit run streamlit_app.py

---

Usage

• Upload a document (PDF, DOCX, image, or text).
• The app extracts and chunks the text, generates embeddings, and stores them in the database.
• Ask questions about the document in the chatbot tab.
• View named entities in both JSON and human-readable formats.
• Use the sidebar to clear all data (for research/testing phase).

---

Deployment

• Deploy on Streamlit Cloud or your own server.
• Set secrets in the Streamlit Cloud UI for API keys and DB credentials.
• Make sure your database is accessible from the deployment environment.

---

Notes

• Research Phase: The "Clear All Data" button wipes the entire database for all users. This is simple for prototyping, but not suitable for production.
• Production Recommendation: For multi-user support, store a user/session ID with each chunk and filter queries/clearing by this ID.
• OCR Model: Requires Mistral API for image OCR.
• LLM Model: Uses Llama 3 via Groq API for Q&A.

---

Future Improvements

• User/session-specific chunk storage and clearing
• User authentication
• More robust error handling and logging
• Support for additional file types and languages

---