Tokenization Interview Questions PDF

Foundational model question on BPE/tokenizers — why token counts differ from words, and production impact on cost, limits, and RAG chunking.

Google's Gemini 1.5 Pro long-context window opens use cases impossible with standard LLMs — whole-codebase analysis, multi-hour video, massive document review. Interviewers test whether you understand real limitations behind the 1M token marketing number.

Cursor rules are context engineering in disguise.

AGENTS.md is becoming the de facto 'README for agents.'

Advanced scenario question on why LLMs hallucinate — next-token prediction vs fact verification, root causes, and production defenses asked at OpenAI, Google, Meta, Anthropic, Microsoft, and Amazon.

Part 2 of the LLM hallucination handbook — factual, citation, reasoning, math, code, temporal, context, and tool-use hallucinations; decoding risks; RAG failure modes; production case study.

Hard Llama/inference question on continuous batching and GPU efficiency.

**vLLM** High throughput continuous batching; popular for multi-tenant APIs.

**Model choice** Flash for simple; Pro for hard multimodal only when needed.

**Use cases** Full codebase questions, hours-long media, multi-doc merger analysis.

**Strategies** Full dump only when eval proves benefit; otherwise RAG + selective long paste.

**Proxy layer** Backend holds keys; assigns budgets per customer/plan.

**Tracing** Span per node: inputs hash, duration, LLM/tool child spans. thread_id correlates user session.

**Stream modes** graph.stream() / astream_events yield incremental updates — node enter/exit, partial state, LLM tokens. Pick mode matching product needs: chat UI wants messages; ops dashboard wants updates.

**Interrupt API** LangGraph can interrupt execution before or after specified nodes — graph pauses, serializes checkpoint (lg-005), and returns control to host app. User sees proposed action; approver decision written to state; graph resumes from checkpoint.

Rate limiting for agents differs from human API usage — bursts are larger and less predictable.

MCP auth interviews test whether you treat agents as privileged automation — because they are.

Embedding similarity, k-NN example banks, MMR diversity, and eval-driven example curation for few-shot prompts.

Compress agent memory into durable distilled state while preserving goals, constraints, and tool outcomes.

Label definitions, calibration examples, abstain classes, and structured outputs for production classifiers without fine-tuning.

Medium AI Agents interview question on agent tracing and replay debugging — architecture, trade-offs, eval, and production patterns.

Medium AI Agents interview question on agent cost caps and budget controllers — architecture, trade-offs, eval, and production patterns.

Hard AI Agents interview question on agent permissions and least privilege — architecture, trade-offs, eval, and production patterns.

Medium AI Agents interview question on stopping criteria and infinite loop prevention — architecture, trade-offs, eval, and production patterns.