Real Interview Questions for the Agentic AI Engineer at Tata Consultancy Services (TCS)
NewPractice MCQs
Answers count toward the Agentic AI Engineer and overall leaderboards; your latest answer to each question is the one that counts.
MCQ-1An HR team's RAG assistant re-ingests policy PDFs nightly. Last Tuesday the leave policy was revised from 20 to 25 days. The logs show the new file was ingested and chunked cleanly, yet answers still quote 20 days, and the vector store holds vectors from both versions. Which change fixes the stale answers?
MCQ-2A TCS-built RAG assistant serves policy documents to employees in three grades. Each chunk carries an `acl_group` value, and the caller's groups come from the JWT. The vector store cannot filter on metadata at query time, so the team can only reindex. A `staff` user asks about executive compensation, and the retriever returns the top 5 chunks by cosine similarity before any filter is applied. Which design keeps restricted chunks out of the staff user's retrieval?
MCQ-3A TCS contract-review assistant indexes each vendor agreement as section-level chunks averaging 6,000 tokens. The deployed model's window is 8,000 tokens, with 1,500 reserved for instructions and the answer, so only two chunks fit. The liability cap for data breaches sits in one section and its carve-out in the next, and the top two chunks hold only the cap. Which change lets the retrieved context include both the cap and the carve-out?
MCQ-4An agent samples the next token on two steps. Step one is peaked: the top token has 0.85 and the next has 0.08. Step two is flat: the top token has 0.14 and the rest of the mass is spread across dozens of tokens. Which setting keeps the candidate count the same on both steps?
MCQ-5An agent calls one LLM to choose among four tools and return JSON arguments, then a second call writes the user-facing summary. At temperature 0.7 the same prompt picked three different tools across 20 regression runs, which broke the test suite, while the summaries should still read naturally. Which setting fixes the tool choices without flattening the summaries?
MCQ-6A support agent exposes two tools: `get_order_status(order_id)` and `cancel_order(order_id, reason)`, described as 'Looks up an order' and 'Cancels an order'. A user asks 'Has order 4471 been cancelled yet?' and the model calls `cancel_order`, which then fails schema validation because `reason` is missing. Which change makes the model pick `get_order_status` for this question?
MCQ-7A TCS claims assistant runs four labelled test queries against a retriever that returns the top 5 chunks, and each query has exactly one relevant chunk. The relevant chunk lands at rank 1, rank 3, outside the top 5, and rank 2 for the four queries respectively. Before generation, the team wants one number that rewards placing the relevant chunk higher. What is the mean reciprocal rank at 5?
MCQ-8A TCS agentic assistant answers policy questions from a retriever over 10,000 chunks. A labelled eval set has 40 questions, each with exactly 3 relevant chunks. For one query the retriever returns K=5 chunks and 2 of them are relevant. The team reports Recall@5 and Precision@5 for this query. Which values are correct, and what does the gap between them tell you about the retrieved set?
MCQ-9A Python RAG loop embeds the question, takes the top 4 chunks by cosine similarity and sends them with the question to the LLM. Each chunk is about 300 tokens and the prompt budget is 4,096 tokens. For 'What is the resignation notice period?' the answer sits in the chunk ranked ninth, and the model replies that the policy does not say. Which change fixes this?
MCQ-10A TCS team builds an assistant for invoice disputes. A single LLM call drafted a reply that cites a refund amount the model never looked up. The team wants the system to query the ledger, then either query the payment record or draft the reply depending on what comes back, so the number of steps varies per dispute. Which design makes this an agent rather than a single-turn generator?
Curated: · Written: · Reviewed: