Top Accenture 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.
Design a RAG system for an enterprise knowledge base.
Connectors to source systems, chunking, embeddings, a vector index with security trimming (users see only permitted docs), retrieval + re-ranking, grounded generation with citations, and an evaluation + monitoring loop.
How do you measure RAG quality?
Retrieval metrics (hit rate, MRR/nDCG) and answer metrics (faithfulness/groundedness, relevance, completeness) on a curated eval set; iterate on chunking, embeddings and re-ranking.
Fine-tuning vs RAG — how do you choose?
RAG for fresh/changing knowledge and citations; fine-tuning for style/format/behaviour or narrow tasks. Often combine: RAG for knowledge, light fine-tune/prompt for behaviour.
Explain an agentic workflow with human-in-the-loop.
An agent plans and calls tools, but pauses at high-risk steps for human approval (via an interrupt/checkpoint), then resumes — balancing autonomy with control and auditability.
How do you defend against prompt injection?
Separate trusted instructions from untrusted content, sanitise/escape retrieved text, constrain tool permissions, validate tool outputs, and add output filtering — never let retrieved content silently override system instructions.
What vector databases have you used and how do you choose?
pgvector (simple, Postgres-native), FAISS (in-memory, fast), Pinecone/Weaviate/Azure AI Search (managed, scalable). Choose on scale, latency, filtering/security and ops overhead.
How do you productionise and monitor an LLM app?
Containerise, add CI/CD, cache, trace with tokens/latency/cost, run offline + online evals, and version prompts — treat the whole pipeline as software.
What makes a strong fresher AI portfolio project?
A deployed RAG assistant or a small agent (e.g. a Jira/GitHub agent) with evaluation and a live URL beats certificates — it proves you can ship.