PrivaDoc AI 🛡️🤖

PrivaDoc AI is a secure, self-hosted Document Retrieval-Augmented Generation (RAG) assistant. It allows teams to upload sensitive documents, automatically parse them with high-fidelity layout analysis, store vector embeddings in a pgvector database, and perform context-aware private Q&A using locally-running Large Language Models (LLMs) via Ollama.

---

🌟 Key Features

• 🔒 Complete Data Privacy: Operates entirely offline using local LLMs (via Ollama) and self-hosted databases. No document content ever leaves your infrastructure.
• 📂 Multi-Modal Document Parsing: Leverages the high-resolution parsing capabilities of the Unstructured API to ingest multi-format files (PDFs, TXT, DOCX, etc.) while preserving structural context (like section titles).
• ⚡ Contextual Vector Search: Chunks documents dynamically and stores their mathematical representations inside PostgreSQL with the pgvector extension using the nomic-embed-text embedding model.
• 👥 Tenant-like Collaboration Groups: Organize documents and chats into Groups. Access control is enforced at the database level so users can only chat with documents in groups they are active members of.
• 💬 History-Aware Q&A: Automatically reformulates follow-up chat messages into standalone queries by analyzing the context of recent conversation history before performing similarity search.
• 🚀 Modern SPA Dashboard: A responsive Angular 21 Single Page Application featuring group managers, member role management, real-time ingestion status indicators, and an interactive chat interface.

---

🏗️ Architecture

PrivaDoc AI uses a modern multi-tier microservices architecture designed to prioritize privacy and scalability:

graph TD
    subgraph Frontend [Client Tier]
        A[Angular 21 SPA]
    end

    subgraph Backend [Application Tier]
        B[Spring Boot App]
        C[Spring Security + JWT]
        D[Spring AI Core]
    end

    subgraph Infrastructure [Data & Inference Tier]
        E[(PostgreSQL + pgvector)]
        F[Unstructured API Service]
        G[Ollama Local Inference]
    end

    A -->|REST API + JWT| B
    B --> C
    B -->|Ingest/Retrieve| E
    B -->|Parse Layouts| F
    D -->|Embeddings / Chat| G

Loading

Component Details

1. Frontend (Angular 21): Captures user interaction, handles upload streams, and renders markdown answers with reference citations.
2. Backend (Spring Boot 3.5): Orchestrates authentication (JWT), file storage, and the RAG pipeline. It leverages Spring AI to interface with Ollama.
3. Vector Database (PostgreSQL + pgvector): Acts as both the relational database (for users, groups, sessions, and messages) and the semantic storage for document vector chunks.
4. Document Parser (Unstructured API): Analyzes document layouts to separate titles, text, tables, and headers for smarter chunking.
5. Local LLM Engine (Ollama): Generates embeddings using nomic-embed-text and processes Q&A answers using local models (default configured: gemma4).

---

🛠️ Local Setup & Run Guide

Follow these steps to run PrivaDoc AI on your local machine.

Prerequisites

Make sure you have the following installed on your machine:

• Docker Desktop
• Java 17 JDK or higher
• Node.js (v20+) & npm
• Ollama

---

Step 1: Start the Local LLM (Ollama)

1. Install and start Ollama on your machine.
2. Download the required embedding and chat models by running the following commands in your terminal:

   ollama pull nomic-embed-text
   ollama pull gemma4

   (Note: You can change the models by updating spring.ai.ollama.chat.options.model and spring.ai.ollama.embedding.options.model inside privadocai/src/main/resources/application.properties.)

---

Step 2: Spin Up the Infrastructure Containers

We use Docker Compose to run PostgreSQL (with pgvector) and the unstructured-api container.

From the root directory of the project, run:

docker-compose up -d postgres-vector unstructured-api

This command starts:

• Postgres Database on port 5432 with username priva-admin, password Ldwpe30idn03@#, and database priva-db.
• Unstructured API on port 8001 (internally port 8000).

---

Step 3: Start the Backend (Spring Boot)

1. Navigate to the backend directory:

   cd privadocai

2. Run the application using Gradle:
   • On Windows (PowerShell/CMD):

     ./gradlew.bat bootRun

   • On Linux/macOS:

     ./gradlew bootRun

The backend server will start running on http://localhost:8080. The database tables and VectorStore schema will be initialized automatically on the first startup.

---

Step 4: Start the Frontend (Angular)

1. Navigate to the client directory:

   cd ../client

2. Install the frontend dependencies:

   npm install

3. Start the Angular development server:

   npm run dev

   (Or run npx ng serve)

Open your browser and navigate to http://localhost:4200 to access the application.

---

🐳 Option: Running the Entire Stack via Docker Compose

If you prefer not to install Node.js and Java locally, you can run the entire application inside Docker containers.

1. Ensure Ollama is running on your host machine and accessible.
2. From the root directory, run:

   docker-compose up --build -d

3. The frontend will be accessible at http://localhost:4200 and the backend api at http://localhost:8080.

---

🔒 Security & Verification Details

• All endpoints under /api/chat/** and /api/document/** require a valid JWT header (Authorization: Bearer <TOKEN>).
• User permissions are validated to ensure users cannot upload, delete, ingest documents, or initiate chats in groups where they are not active members.
• Document files are stored locally in the directory specified by app.storage.upload-dir (default: ./documents).