🚀 TechPrep AI — Technical Interview Knowledge Base

A production-grade RAG system built for technical interview preparation. Query a curated knowledge base covering System Design, DSA, and AI/ML concepts using natural language.

---

📌 What Is This?

As a final year CS student preparing for AI engineering roles, I needed a single tool that could answer technical interview questions across System Design, DSA, and AI/ML — grounded in real curated content, not hallucinated answers.

So I built TechPrep AI — a production-grade RAG pipeline that:

• Retrieves relevant content using Hybrid Search
• Reranks results using Cohere Cross-Encoder
• Generates answers using Groq LLM
• Tracks every query using Langfuse Observability
• Evaluates quality using a Custom Eval Pipeline

---

🏗️ Architecture

User Query │ ▼ ┌─────────────────────────────────┐ │ Hybrid Retrieval │ │ BM25 Search + Dense Search │ │ Reciprocal Rank Fusion (RRF) │ │ Cohere Cross-Encoder Reranking │ └─────────────────────────────────┘ │ ▼ ┌─────────────────────────────────┐ │ LLM Generation │ │ Groq API (Llama 3.1 8B) │ │ Context-grounded prompt │ └─────────────────────────────────┘ │ ▼ ┌─────────────────────────────────┐ │ Observability │ │ Langfuse Tracing │ │ Latency + Cost Tracking │ │ User Feedback Logging │ └─────────────────────────────────┘

---

📊 Evaluation Results

Metric	Score
Answer Relevancy	0.84 / 1.0
Faithfulness	0.24 / 1.0 (improving with better data)
Avg Query Latency	~6000ms
Retrieval Latency	~5500ms
Generation Latency	~800ms

Evaluated on 5 golden questions using custom LLM-as-judge eval pipeline powered by Groq.

---

⚙️ Tech Stack

Layer	Technology
LLM	Groq API — Llama 3.1 8B / 70B
Vector DB	Qdrant Cloud
Embeddings	FastEmbed
Retrieval	BM25 + Dense + RRF
Reranking	Cohere Rerank v3
Observability	Langfuse
Backend	FastAPI
Frontend	Streamlit
Language	Python 3.12

---

🔍 Key Features

1. Hybrid Search

Combines BM25 keyword search with dense vector search. BM25 catches exact keyword matches while dense search captures semantic meaning. Results fused using Reciprocal Rank Fusion (RRF).

2. Cohere Cross-Encoder Reranking

After RRF fusion, top candidates are reranked using Cohere's cross-encoder model for higher precision retrieval. This significantly improves answer quality over basic vector search alone.

3. Langfuse Observability

Every query is traced end-to-end with:

• Retrieval latency
• Generation latency
• Total latency
• User feedback (helpful / not helpful)

4. Custom Eval Pipeline

LLM-as-judge evaluation pipeline using Groq that scores:

• Faithfulness — is the answer grounded in context?
• Answer Relevancy — does it answer the question?

5. PDF Support

Supports .txt, .md, and .pdf document ingestion — upload your own notes or textbooks.

---

📁 Project Structure

techprep-ai/ ├── app/ │ ├── config.py # All config and env vars │ ├── loaders.py # Document loaders (txt, md, pdf) │ ├── chunkers.py # Text chunking with metadata │ ├── ingest.py # Embedding and Qdrant upload │ ├── retrieval.py # Hybrid search + reranking │ ├── query.py # RAG pipeline + tracing │ ├── eval.py # Evaluation pipeline │ ├── llm.py # LLM clients │ └── tracer.py # Langfuse tracing ├── data/ │ └── raw/ # Knowledge base documents ├── main.py # FastAPI backend ├── dashboard.py # Streamlit frontend ├── requirements.txt └── .env

---

🚀 How To Run

1. Clone the repo

git clone https://github.com/yourusername/techprep-ai
cd techprep-ai

2. Create virtual environment

python -m venv venv
venv\Scripts\Activate.ps1  # Windows

3. Install dependencies

pip install -r requirements.txt

4. Set up environment variables

Create a .env file:

GROQ_API_KEY=your_key QDRANT_URL=your_url QDRANT_API_KEY=your_key COHERE_API_KEY=your_key LANGFUSE_PUBLIC_KEY=your_key LANGFUSE_SECRET_KEY=your_key

5. Add your documents

Put .txt, .md, or .pdf files in data/raw/

6. Ingest documents

python -m app.ingest

7. Run the app

# Terminal 1
uvicorn main:app --reload

# Terminal 2
streamlit run ui.py

---

💡 Why I Built This

As a final year CS student actively interviewing for AI engineering roles, I found no single tool that could answer technical questions across System Design, DSA, and AI/ML in a grounded, reliable way.

I built TechPrep AI to solve my own problem — and used it to prepare for interviews. The production-grade architecture (hybrid retrieval, reranking, observability, eval) reflects how real AI systems are built.

---

🔮 Future Improvements

• Add more high quality domain data
• Implement query caching for faster responses
• Add conversation memory for follow-up questions
• Deploy on cloud (Railway / Render)
• Add voice input support

---

🌐 Live Demo

• Frontend: https://prepai-h8ekghmvjxyfapdgbc9dth.streamlit.app/
• Backend API: https://web-production-5af08.up.railway.app/docs

👩‍💻 Author

Srashti Final Year IT Student | Aspiring AI Engineer

• GitHub: https://github.com/SRASHTI2004
• LinkedIn: https://www.linkedin.com/in/srashti-choudhary
• Email: srashtichoudhary5@gmail.com