InterviewAssistant – Grounded RAG-Based Resume & Interview System

A retrieval-augmented generation system that helps job seekers tailor resumes, draft grounded cover letters, and generate role-specific interview questions using structured candidate knowledge aligned to a job description.

Unlike generic LLM prompts, this system separates retrieval from generation to prevent hallucinated experience and ensure claims are backed by actual candidate evidence.

---

🚀 What This Solves

Generic LLMs:

• Blend job description requirements with invented experience
• Hallucinate tools, metrics, or skills
• Produce generic cover letters

This system:

• Stores structured candidate experience in a persistent FAISS vector store
• Builds an ephemeral index of each uploaded job description
• Performs two-stage retrieval (JD → Candidate)
• Applies strict grounding guardrails before LLM generation

---

🏗 Architecture

User Uploads Job Description
↓
Ephemeral JD Vector Index (FAISS)
↓
Retrieve Key JD Requirements
↓
Retrieve Matching Candidate Evidence (Persistent FAISS)
↓
Construct Grounded Prompt with Evidence Sections
↓
Groq LLaMA 3.1 8B (API)
↓
Interview Questions / Cover Letter

---

🧠 Core Design Decisions

Persistent Candidate Knowledge Base

• Resume, work experience, projects, STAR stories
• Embedded using sentence-transformers
• Stored in FAISS
• Rebuilt only when candidate docs change

Ephemeral JD Index

• Built per uploaded JD
• Ensures role-specific requirement extraction
• Prevents cross-role contamination

Two-Stage Retrieval

1. Retrieve important JD requirement chunks
2. Retrieve aligned candidate evidence

This prevents hallucinated claims and enforces grounding.

Grounding Guardrails

• Explicit system rules
• Structured evidence formatting
• No invention of unsupported experience
• Gap awareness if JD requirement lacks supporting evidence

API-Based Generation

• Groq LLaMA 3.1 8B
• Fast inference
• Reduced local resource usage
• Controlled output token limits for cost efficiency

---

🛠 Tech Stack

• Python
• Streamlit
• LangChain
• FAISS (local vector store)
• SentenceTransformers (embeddings)
• Groq API (LLaMA 3.1 8B)
• python-dotenv

---

⚙️ Setup

1. Clone repository

git clone https://github.com/99anjalipai/InterviewAssistant.git
cd InterviewAssistant

2. Create virtual environment

python -m venv rag_env
rag_env\Scripts\activate

3. Install dependencies

pip install -r requirements.txt

4. Add GROQ API Key

Create a .env file in the project root:

GROQ_API_KEY=your_api_key_here

5. Build Candidate Vector Store

Place your candidate documents (resume, work experience, projects, STAR stories) inside:

candidate_docs/

Then run:

python -m ingest.build_candidate_index

This process will:

• Chunk your documents
• Generate embeddings using sentence-transformers
• Create a persistent FAISS index inside:

stores/candidate_faiss/

6. Run the Application

streamlit run app.py

Open the provided local URL in your browser to start using the application.

📁 Project Structure

• app.py – Streamlit frontend
• rag_engine.py – Retrieval and grounding pipeline
• scripts/llm_groq.py – Groq LLM wrapper
• prompts.py – Prompt templates for generation
• ingest/ – Candidate index builder
• candidate_docs/ – Private candidate knowledge base (gitignored)
• stores/ – FAISS vector store (gitignored)

📈 Future Improvements

• Resume diff highlighting against JD
• Skill coverage scoring
• Explicit gap analysis report
• Multi-JD comparison
• Conversational interview simulation mode
• Deployment version (Docker or Streamlit Cloud)

🔒 Notes

• Candidate documents and vector stores are excluded from version control.
• API usage is token-limited to control cost.
• Retrieval parameters are tuned for grounding accuracy and efficiency.