Real-world questions based on production deployments, not textbook theory.
Based on production AI systems deployed across MENA enterprises

---

Difficulty Levels

	Level	Target	Description
🟢	Fresh	Entry-level	Core concepts, fundamentals, "what is X" questions
🟡	Intermediate	1-3 years	Implementation details, trade-offs, "how would you" questions
🔴	Advanced	3-5 years	Production exposure, debugging, architecture decisions
⚫	Expert	5+ years	System design, scale challenges, strategic decisions

---

Sections

13 Questions · Architecture, Self-RAG, Agentic RAG, GraphRAG, Multimodal RAG 🟢🟡🔴⚫	10 Questions · Fixed, recursive, semantic, late chunking, LLM-based 🟢🟡🔴⚫	8 Questions · HNSW, IVF, PQ, multi-tenancy, bias 🟢🟡🔴⚫	6 Questions · BM25, RRF, reranking, cross-encoders 🟡🔴⚫
6 Questions · Cache strategies, invalidation, cost savings 🟢🟡🔴⚫	9 Questions · Orchestrator, ReAct, coordination tax 🟢🟡🔴⚫	6 Questions · Tool use, schema design, error handling 🟢🟡🔴⚫	8 Questions · Tokenization, dialects, embeddings, RTL 🟢🟡🔴⚫
8 Questions · Quantization, vLLM, batching, GPU memory 🟢🟡🔴⚫	6 Questions · LoRA, QLoRA, data quality, evaluation 🟢🟡🔴⚫	6 Questions · RAGAS, LLM-as-Judge, golden datasets 🟢🟡🔴⚫	6 Questions · Prompt injection, OWASP, PII, defense layers 🟢🟡🔴⚫
5 Questions · Token costs, caching, model routing 🟡🔴⚫	4 Questions · Tracing, metrics, alerting, dashboards 🟡🔴⚫	5 Questions · Hallucinations, CoT failures, edge cases 🟡🔴⚫	5 Questions · End-to-end architectures, scaling, trade-offs ⚫

---

Architecture Diagrams

Each section includes detailed architecture diagrams. Here's a preview:

RAG Pipeline Architecture	Chunking Strategies	Hybrid Search Pipeline
Multi-Agent Systems	Semantic Caching	Vector DB Architecture

---

How to Use This Guide

For Interview Prep Start with sections matching your experience level Read the question first, try to answer it yourself Compare with the expected answer Pay attention to Red Flags — interviewers watch for these Practice explaining concepts out loud

For Team Training Use as a structured knowledge assessment Assign sections based on team roles Discuss answers in group sessions Build internal knowledge base from answers Track improvement over time

---

Quick Reference

Key Stats from Production Deployments

Insight	Source
90% of agentic RAG projects failed in production in 2024	RAG Systems
Semantic chunking improved faithfulness from 0.47 to 0.79	Chunking
1B vectors at 1024 dims = ~4TB storage	Vector DBs
Hybrid search: +15-30% precision over vector-only	Hybrid Search
Semantic caching: $52K → $4.8K monthly (90% reduction)	Caching
Optimal multi-agent count is typically 3-4 agents	Multi-Agent
Arabic tokenizers use 3-5x more tokens than English	Arabic NLP

---

Bonus Sections

🚩 Red Flags Common mistakes that reveal shallow understanding during interviews. Categorized by topic: RAG, Multi-Agent, Deployment, Evaluation, Security.

📚 Resources Curated papers, frameworks, and evaluation tools. Including landmark papers on RAG, Self-RAG, and the Google DeepMind agent scaling study.

---

Contributing

Contributions welcome! Please:

1. Open an issue for discussion first
2. Submit a PR with new questions
3. Include difficulty level and category
4. Add expected answer and red flags
5. Follow the existing format

---

License

MIT License — Feel free to use for interview prep, team training, or educational purposes.

---

Built with real-world experience from production AI systems in the MENA region.

Last updated: February 2026