Top TCS AI Interview Questions — RAG & Agentic AI (2026) with clear, practical answers — curated for 2026 interviews. Pair these with a deployed project from our AI, GenAI & Agentic AI training.
Walk through a production RAG pipeline end to end.
Ingest → chunk (size/overlap tuned to content) → embed → store in a vector DB → at query time embed the question, retrieve top-k, optionally re-rank, build a grounded prompt with citations, generate, and evaluate faithfulness. Add caching and guardrails for production.
How do you reduce hallucination in RAG?
Ground strictly in retrieved context, cite sources, add "say I don’t know" instructions, re-rank for relevance, and run groundedness evaluation. Better retrieval usually beats a bigger model.
What chunking strategy do you use and why?
Semantic/structure-aware chunks (by heading/paragraph) with small overlap, sized to the embedding model and the question type. Too large dilutes retrieval; too small loses context.
What is an AI agent and when is it justified over RAG?
An agent uses an LLM to plan and call tools in a loop. Justify it when the task needs multiple steps/tools or decisions; for simple Q&A over docs, RAG alone is cheaper and more reliable.
What is MCP (Model Context Protocol)?
A standard protocol for exposing tools/resources to AI agents, so an agent can act on systems (Jira, GitHub, databases) through a consistent interface with controlled scope and auth.
How do you evaluate an agentic system?
Task success rate, step efficiency, tool-call correctness, and failure analysis — on a fixed eval set, with human-in-the-loop for high-risk actions.
How do you control cost and latency?
Smaller/cheaper models where possible, caching, limiting retrieved tokens and agent steps, streaming responses, and routing easy queries to lighter paths.
How would you keep an enterprise AI assistant secure?
Access-control-trimmed retrieval, PII handling, prompt-injection defences, audit logging, and human approval for sensitive actions.