Core vocabulary
Practice this questionWANT TO UNDERSTAND THE CONCEPT FIRST?Learn Core vocabulary
- Use this after studying the full explanations.
- Spend 10 minutes on vocabulary, 10 on RAG and agents, 10 on security/workflows/evaluation, and 10–30 on the insurance design and scenario answers.
| Concept | Say this in an interview | Remember the limit |
|---|---|---|
| LLM | A trained model that processes/generates language using learned patterns and supplied context | It does not automatically know or verify private account facts |
| Tokens | Model processing units used for context limits, throughput, and often billing | Words/characters only estimate usage |
| Context | Information supplied for the current model call | Large capacity is not perfect attention or permanent memory |
| Prompting | A clear task contract with inputs, constraints, output, and unknown behavior | Prompts do not enforce backend permissions |
| Zero-shot / few-shot | Instructions alone / instructions plus examples | Evaluate example quality and token cost |
| Structured output | Model data constrained to a declared schema where supported | Valid shape is not factual or business validity |
| Tool calling | Model proposes a named operation and arguments; application validates and executes | Never grant direct broad database or network authority |
| Embedding | A vector representing useful content features for similarity | Similar meaning does not prove truth, access, or applicability |
| Cosine | Similarity of vector direction | Score is not probability of correctness |
| Vector DB/index | Storage/indexing for similarity retrieval and metadata | Filtered recall and lifecycle matter |
| BM25 | Lexical relevance ranking over matching terms | Good for identifiers; paraphrases may need semantic retrieval |
| Hybrid search | Combine lexical and vector candidates | Fuse ranks or calibrated scores, not arbitrary score addition |
| Reranking | Carefully order a retrieved candidate set | Cannot recover missing candidates |
Adapted from 13_AI.MD · “Core vocabulary”. Code examples are study fragments. Difficulty and core labels are editorial estimates; version-specific behavior should be checked against your project.