Resume Q&A System with RAG, HyDE and Multi-Turn Chat

A semantic question-answering system built on a personal resume, using RAG, Hypothetical Document Embeddings (HyDE), FAISS dense indexing, BM25 sparse retrieval, and GMI Cloud LLM API.

Features

• Hybrid retrieval strategy combining:
  • Dense semantic search (FAISS + MiniLM)
  • Sparse keyword search (BM25)
• Hypothetical document expansion (HyDE) to boost recall
• Structured prompt construction including:
  • Retrieved resume snippets
  • Optional HyDE document
  • Multi-turn chat history
• Asynchronous LLM calls via GMI Cloud (DeepSeek-R1)
• Command-line interface supporting multi-turn conversations
• Benchmark script for comparing sequential vs. concurrent queries

Project Structure

rag_demo/
├── data/                   
│   └── resume.pdf           # Input resume file (PDF, TXT, or MD)
├── embeddings/             
│   ├── resume_embeddings.npy
│   └── texts.json           # Mapping from vectors back to text
├── index/                  
│   └── resume.index         # FAISS index file
├── src/
│   ├── preprocess.py        # Load and semantically split resume
│   ├── embed.py             # Generate embeddings from text chunks
│   ├── build_index.py       # Build and persist FAISS index
│   ├── query_engine.py      # Core RAG + HyDE logic and GMI API integration
│   ├── benchmark.py         # Compare sequential vs. asynchronous query performance
│   └── evaluation.py        # Retrieval and generation evaluation scripts
└── .env                     # Your GMI Cloud API key (GMI_API_KEY=…)

Installation

1. Clone the repository and create a virtual environment

   git clone https://github.com/yourusername/rag_demo.git
   cd rag_demo
   python3 -m venv venv
   source venv/bin/activate
   
   

2. Install dependencies

   pip install -r requirements.txt

3. Create a .env file in the project root:

   GMI_API_KEY=your_api_key_here
   

Usage

1. Preprocess the resume

python src/preprocess.py

2. Generate embeddings

python src/embed.py

3. Build the FAISS index

python src/build_index.py

4. Run the multi-turn CLI

python src/query_engine.py

5. Run the benchmark

python src/benchmark.py

Model and API Details

Embedding model: all-MiniLM-L6-v2 via sentence-transformers

LLM model: deepseek-ai/DeepSeek-R1 on GMI Cloud

API endpoint: https://inference-engine.gmicloud.ai/v1/chat/completions

Result