AI Engineer (LLM & Generative AI) question bank
100 questions. Filter by level and skill, track what you've solved, and see how many people get each one right.
- What is a token
- What text embeddings are
- Effect of temperature
- What the context window limits
- What RAG does
- What a hallucination is
- Purpose of a system prompt
- Comparing two embeddings
- Dot product of two vectors
- Length of a vector
- Softmax of equal scores
- What softmax outputs look like
- Temperature close to zero
- What top-k sampling does
- Grouping data with no labels
- Why keep a separate test set
- Spotting overfitting
- Computing precision
- Why activation functions exist
- A learning rate that is too high
- How rare words get tokenized
- What LLM pretraining learns
- What few-shot prompting is
- Asking the model to reason first
- The required keyword in JSON Schema
- What a tool definition contains
- The basic agent loop
- What MCP is for
- What a vector database does
- Which words BM25 rewards
- What a golden set is for
- Handling an HTTP 429 response
- Customer data and a hosted LLM
- Parsing a JSON reply in Python
- Choosing a chunking strategy
- Getting reliable JSON output
- Knowledge that changes every day
- Embeddings miss exact product codes
- Adding a reranker
- Who runs the tool
- Making prompt caching work
- Pairwise LLM judge bias
- Cosine similarity by hand
- Temperature changes probabilities
- How many tokens top-p keeps
- From logprobs to probability
- Computing recall at 5
- Computing MRR
- Estimating monthly token cost
- Savings from a cached prefix
- Why add jitter to backoff
- Tracing a retry helper
- What schema-constrained output guarantees
- Two tool calls in one turn
- Counting chunks with overlap
- Keeping tenants apart in RAG
- Tuning HNSW search breadth
- Reciprocal rank fusion scores
- Judge prefers longer answers
- Dot product on unit vectors
- Cosine of a scaled vector
- Streaming tool-call arguments
- Semantic cache gives wrong answers
- Debugging a bad agent answer
- Which rate limit binds first
- Memory for 4-bit weights
- Running SQL that an LLM wrote
- Users start asking new things
- Injected instructions in a web page
- Why long contexts eat GPU memory
- Stuffing 100 chunks into the prompt
- What LoRA trains
- Shipping a prompt change safely
- Agent stuck in a loop
- Upgrading the embedding model
- What streaming improves
- Sizing the KV cache
- Why attention divides by sqrt(d_k)
- What the causal mask prevents
- Counting LoRA parameters
- Where generation time goes
- Estimating response time
- Temperature 0 is not deterministic
- Counting letters in a word
- Data leaking through image links
- A judge with 90% agreement
- Computing nDCG at 3
- Is a 4-point gain real
- When a tool call fails
- A suspicious third-party MCP server
- Budgeting the context window
- How speculative decoding helps
- Continuous batching in serving
- Serving a mixture-of-experts model
- How many results to rerank
- Adding BM25 and cosine scores
- Fine-tuned model scores too well
- Field order in an output schema
- One failure in Promise.all
- Jailbreaks spread across turns