SHL Assessment Recommendation System

This project implements a simple LLM-based, retrieval-augmented recommendation system for the SHL assignment.

The system recommends relevant SHL Individual Test Solutions based on:

• Natural language queries
• Job description text
• Job description URLs

---

🔗 Project Links

• Web App URL: https://assessmentrecommendation-y3wxlazwen8zz7zythiq8n.streamlit.app/
• API URL: https://shl-render.onrender.com/recommend
• GitHub Repo: https://github.com/555vedant/Assessment_Recommendation

---

System Architecture

---

How It Works

1. SHL product catalog is scraped and stored locally
2. Assessments are converted into embeddings
3. FAISS is used as a vector database
4. An LLM is used for query understanding
5. Relevant assessments are retrieved and ranked
6. Balanced recommendations are returned

---

Deployment Constraints & Design Decisions (Important)

During deployment, the system was initially configured with SentenceTransformer embeddings + FAISS inside the live API.

However, the Render free tier (512 MB RAM limit) consistently caused out-of-memory crashes when loading:

• Embedding models
• Vector indices
• Torch runtime dependencies

CSV Generation (Evaluation Mode)

The system generates a CSV file named:

predictions.csv

Format:

Query,Assessment_url

This CSV is used for automated evaluation using Mean Recall@10.

---

Assignment Compliance

• Scraped SHL Individual Test Solutions
• LLM-based query understanding
• Retrieval-augmented pipeline
• Balanced recommendations
• Working web application
• Correct CSV format for submission

---

Author

Vedant Kasar