Corporate Policy & Knowledge Assistant (RAG)

1. Project Summary

A full-stack Retrieval-Augmented Generation (RAG) application designed to streamline access to corporate knowledge. The system allows users to upload internal documents (PDFs, policies) and query them using natural language. Built with Spring Boot and Spring AI on the backend, it leverages PostgreSQL with pgvector for efficient vector similarity search. The frontend is a modern, responsive React application featuring a dark-mode chat interface. This project demonstrates the practical application of Generative AI to solve enterprise information retrieval challenges, ensuring accurate, context-aware responses by grounding LLM outputs in proprietary data.

2. Key Features & Highlights

• Spring AI Integration: Seamlessly orchestrates LLM interactions and vector store operations.
• RAG Architecture: Retrieves relevant context to ground LLM responses in facts.
• Vector Database: Uses PostgreSQL with pgvector for high-performance similarity search.
• Document Ingestion: Automated parsing and chunking of multi-format documents via Apache Tika.
• Modern Frontend: React-based chat interface with real-time streaming-like experience.
• Scalable Design: Containerized database and stateless backend services for easy scaling.
• Secure Data Handling: Processes sensitive corporate data locally before vectorization.
• Context-Aware: Retrieves top-k relevant document chunks for precise answer generation.

3. Detailed Technical Analysis & Flow Breakdown

System Architecture

The application follows a classic 3-tier architecture enhanced with AI capabilities.

Detailed Data Flow

A. Document Ingestion Pipeline

This process transforms raw files into searchable vector embeddings.

1. Upload: User uploads a file (PDF, DOCX, TXT) via the React frontend.
2. API Endpoint: The file is sent to POST /api/ingest.
3. Reading: TikaDocumentReader extracts text content from the binary file format.
4. Splitting: TokenTextSplitter breaks the continuous text into smaller, manageable chunks (tokens) to fit within the LLM's context window and improve retrieval precision.
5. Embedding & Storage: These chunks are converted into vector embeddings and stored in the vector_store table in PostgreSQL.

sequenceDiagram
    participant User
    participant Frontend
    participant Backend
    participant DB as VectorStore (PG)

    User->>Frontend: Upload Document
    Frontend->>Backend: POST /api/ingest (MultipartFile)
    Backend->>Backend: Parse File (Apache Tika)
    Backend->>Backend: Split Text (TokenTextSplitter)
    Backend->>DB: Store Embeddings
    Backend-->>Frontend: 200 OK "Ingestion Successful"
    Frontend-->>User: Show Success Message

Loading

B. Retrieval & Generation (Chat) Flow

This is the core RAG loop that answers user questions.

1. User Query: User types a question (e.g., "What is the remote work policy?").
2. Similarity Search: The backend converts the query into a vector and searches the PostgreSQL database for the top 3 most similar document chunks (topK=3).
3. Prompt Engineering: A system prompt is dynamically constructed. It instructs the LLM to answer only using the provided context.
   • Template: "You are a helpful assistant. Use the following information to answer the question..."
   • Context: The text content from the retrieved database chunks.
4. LLM Inference: The constructed prompt (System Instruction + User Query) is sent to the Chat Client (e.g., OpenAI API).
5. Response: The LLM generates a natural language answer based on the retrieved facts, which is returned to the frontend.

sequenceDiagram
    participant User
    participant Frontend
    participant Backend
    participant DB as VectorStore (PG)
    participant LLM as AI Model

    User->>Frontend: "What is the remote work policy?"
    Frontend->>Backend: POST /api/chat {query}
    
    rect rgb(240, 248, 255)
        note right of Backend: Retrieval Phase
        Backend->>DB: Similarity Search (Query Vector)
        DB-->>Backend: Return Top 3 Relevant Chunks
    end
    
    rect rgb(255, 240, 245)
        note right of Backend: Generation Phase
        Backend->>Backend: Construct Prompt (Context + Query)
        Backend->>LLM: Send Prompt
        LLM-->>Backend: Generated Answer
    end
    
    Backend-->>Frontend: JSON {answer: "..."}
    Frontend-->>User: Display Answer

Loading

Technology Stack Details

Component	Technology	Purpose
Frontend	React 19, Vite	Fast, reactive user interface.
Styling	Tailwind CSS	Modern, utility-first styling for a premium look.
Backend Framework	Spring Boot 3.2	Robust Java-based REST API.
AI Orchestration	Spring AI	Abstraction layer for switching AI providers and managing vectors.
Vector Database	PostgreSQL + pgvector	Storing and querying high-dimensional vector data.
Document Parsing	Apache Tika	Extracting text from various file formats (PDF, DOC, etc.).
Containerization	Docker Compose	Orchestrating the database and application services.