System Architecture

RAG Pipeline Flow

flowchart TB
    subgraph Client["User Interface"]
        UI[Streamlit Application]
    end
    subgraph Processing["CV Processing Pipeline"]
        Upload[CV Upload]
        Validate[Validation & Format Detection]
        Parse[Text Extraction]
        Chunk[Chunking - 1000 chars with 200 overlap]
        Embed[OpenAI Embeddings]
    end
    subgraph Storage["Local Storage"]
        CV[(CV Document Store)]
        Index[FAISS Vector Index\n1536 dimensions\nCosine similarity]
        Memory[Conversation History]
    end
    subgraph RAG["Interview Q&A Pipeline"]
        Query[Interview Question]
        QueryEmbed[Query Embedding]
        Search[Similarity Search]
        Context[CV Context Retrieval]
        Prompt[Prompt with CV Context]
        LLM[GPT-4 Response]
    end
    UI --> Upload
    Upload --> Validate
    Validate --> Parse
    Parse --> Chunk
    Chunk --> Embed
    Embed --> Index
    Parse --> CV
    UI --> Query
    Query --> QueryEmbed
    QueryEmbed --> Search
    Search --> Index
    Index --> Context
    Context --> Prompt
    Memory --> Prompt
    Prompt --> LLM
    LLM --> UI
    LLM --> Memory
    style Index fill:#8B7355,color:#fff
    style LLM fill:#F5E6D3,color:#333

Technical Flow

Document Processing (One-time):

1. CV Upload (PDF, DOCX, TXT, MD)
2. Text Extraction → Raw CV content
3. Chunking → 1000-character segments with 200-char overlap
4. Embedding → Convert to 1536-dimensional vectors
5. FAISS Indexing → Fast similarity search

Query Pipeline (Per Question):

1. User asks interview question
2. Question embedded to vector
3. FAISS searches for 3 most similar CV chunks
4. Chunks + question + conversation history → GPT-4
5. GPT-4 generates answer based on CV content
6. Response displayed with source context

Installation

Prerequisites

• Python 3.8+
• OpenAI API key
• Streamlit
• LangChain
• FAISS
• OpenAI Python SDK

Steps

1. Clone the repository
2. Create and activate a virtual environment (python -m venv .venv)
3. Install dependencies (pip install -r requirements.txt)
4. Configure OpenAI API key in .env
5. Run streamlit run app.py

Configuration

Edit config.py for:

• FAISS index storage path
• Embedding model: text-embedding-3-small
• LLM: GPT-4
• Chunk size (1000), overlap (200)
• Allowed file types (TXT, PDF, DOCX, MD)
• Maximum file size (50MB)

Key settings:

CHUNK_SIZE = 1000           # Characters per chunk
CHUNK_OVERLAP = 200         # Overlap between chunks
MODEL = "gpt-4"             # OpenAI model
TEMPERATURE = 0.7           # Response creativity
SEARCH_K = 3                # Number of chunks to retrieve
MAX_FILE_SIZE = 50 * 1024 * 1024  # 50MB limit
ALLOWED_EXTENSIONS = ["txt", "pdf", "docx", "md"]

Knowledge Base

The system supports uploading your CV in various formats:

How It Works

1. Upload Your CV: Use the sidebar to upload your resume (one file recommended)
2. Automatic Processing: System extracts text, chunks into segments, and creates searchable index
3. Ask Questions: Query your CV like talking to a recruiter

Typical Interview Questions You Can Ask

Technical Skills:

• "What programming languages do I know?"
• "What are my technical skills?"
• "Do I have experience with [specific technology]?"

Work Experience:

• "Tell me about my work experience"
• "What was my role at [company name]?"
• "How many years of experience do I have in [field]?"

Projects:

• "What are my most relevant projects?"
• "Describe my [project name] project"
• "What technologies did I use in my projects?"

Education:

• "What is my educational background?"
• "Where did I study?"
• "What degrees do I have?"

Achievements:

• "What are my key achievements?"
• "What awards or recognition have I received?"

Usage

Quick Start Guide

1. Launch Application

streamlit run app.py

2. Upload Your CV

   • Click "Knowledge Base Management" in sidebar
   • Upload your CV (PDF, DOCX, TXT, or MD)
   • Wait for processing confirmation

3. Start Interview Prep

   • Type questions in the chat interface
   • Get instant answers based on your CV
   • Follow up with clarifying questions
   • Conversation memory maintains context

Example Conversation

You: What are my technical skills?
Assistant: Based on your CV, your technical skills include Python, JavaScript, 
React, Node.js, SQL, MongoDB, Docker, and AWS. You also have experience with 
Git and Agile methodologies.

You: How many years of Python experience do I have?
Assistant: According to your CV, you have 5 years of Python experience, 
starting from your role as Junior Developer at TechCorp in 2020 through 
your current position.

You: What's my most impressive project?
Assistant: Your most impressive project appears to be the E-commerce Platform 
you built at Current Company, which handled 10,000+ daily transactions and 
reduced page load times by 40% using React and Node.js microservices.

Managing Your Session

• Clear Chat: Reset conversation to start fresh preparation
• Upload New CV: Replace with updated resume (rebuilds index automatically)
• Multiple Questions: Ask follow-ups - system remembers previous context

Deployment

Streamlit Community Cloud

1. Push code to GitHub
2. Connect repository at share.streamlit.io
3. Configure secrets (OPENAI_API_KEY)
4. Deploy application

Production Considerations

• Secure API keys in environment variables
• Monitor OpenAI API usage and costs
• Implement rate limiting for queries
• Regular vector store backups
• User authentication for internal systems

Project Structure

Rag/
│
├── app.py                        # Main Streamlit application
├── rag_engine.py                 # RAG pipeline with FAISS
├── config.py                     # Configuration management
├── utils.py                      # File processing utilities
├── requirements.txt              # Python dependencies
├── .env                          # Environment variables (not in repo)
├── .gitignore                    # Git ignore rules
└── README.md                     # This file
│
├── data/
│   ├── documents/               # CV storage (created on first upload)
│   └── faiss_store/             # FAISS vector index (auto-generated)
│       ├── index.faiss          # Vector index file
│       └── index.pkl            # Metadata pickle
│
└── assets/                       # Static assets

Technical Documentation

RAG Pipeline Details

• Document Processing: Multi-format support with atomic file operations
• Text Chunking: 1000 characters with 200-character overlap
• Embedding Generation: OpenAI text-embedding-3-small (1536 dimensions)
• Vector Search: FAISS with cosine similarity
• Context Retrieval: Top-3 most relevant chunks
• Response Generation: GPT-4 with conversation memory

Acknowledgments

Developed for: Silvertree Brands - Interview Preparation Tool Challenge
Development Time: <4 hours (Rapid Prototype)
Technical Documentation: Confluence Wiki

Technology Partners:

• OpenAI for GPT-4 and embeddings
• Facebook AI Research for FAISS
• LangChain for RAG framework
• Streamlit for rapid UI development

Key Learning: Combining retrieval precision with LLM reasoning eliminates hallucinations while maintaining natural conversation flow - perfect for interview preparation where accuracy matters.

• No rate limiting (vulnerable to abuse)
• OpenAI API key exposed to server (not user-isolated)

📚 Technical Stack

Core Dependencies