Claude Certified Architect - Foundations flashcards

32 cards. Read the term and the definition here. The deck at the top of this page is optional.

Context window

The maximum number of tokens a single request may contain across system prompt, history, tool definitions, retrieved documents, and generated output. Exceeding it forces truncation or summarization.

Temperature parameter

A sampling setting that scales randomness. Values near zero make output nearly deterministic and repeatable; higher values broaden the distribution and increase variety and variance.

Extended thinking

A mode where the model produces additional internal reasoning tokens before its final answer. It raises latency and cost but improves accuracy on multi-step problems.

Training cutoff

The date after which the model has no native knowledge. Facts past this point must be supplied through retrieval or directly in the prompt.

Stop sequence

A string that halts generation as soon as the model produces it. It truncates output at a chosen boundary rather than summarizing or restructuring it.

System prompt role

It establishes persistent persona, constraints, and output format for the whole conversation. Task-specific user input should not be trusted to override it.

Few-shot prompting

Including several worked input-output examples in the prompt. Examples steer format and reasoning style more reliably than descriptive adjectives, and precede the target input.

Chain-of-thought prompting

Asking the model to produce intermediate reasoning steps before the final answer. It measurably improves accuracy on multi-step arithmetic, logic, and planning tasks.

Assistant prefill

Supplying the opening text of the model's reply so generation continues from that prefix. It constrains output format without adding extra instructions.

Delimiter isolation

Wrapping untrusted content in tags or quotes so it is clearly separated from instructions. This reduces instruction bleed and prompt-injection spillover.

Tool use loop

The cycle where the model emits a tool call, the client executes it and returns the result, and the model continues until it stops requesting tools.

Tool input schema

A JSON Schema describing a tool's accepted arguments. The model uses it to construct calls; validation failures should be returned as error results for correction.

Iteration budget

A cap on the number of agent turns or tool calls per run. It bounds cost and latency and prevents runaway loops from a misbehaving agent.

Parallel tool calls

Requesting multiple tool invocations in one turn. Valid only when calls are independent; dependent calls must be sequenced so later arguments can use earlier results.

Human-in-the-loop gate

A required human approval before executing an irreversible or high-impact action such as a payment, deletion, or external write. Read-only lookups usually skip it.

RAG pipeline stages

Chunk documents, embed them, index vectors, retrieve top-k by similarity, optionally rerank, then place selected passages in the prompt for grounded generation.

Chunk size tradeoff

Large chunks dilute embeddings and waste tokens; small chunks lose surrounding context. Chunk boundaries should follow semantic units, not fixed character counts alone.

Hybrid retrieval

Combining dense vector similarity with lexical matching such as BM25. Dense handles paraphrase and semantics; lexical handles exact names, codes, and rare terms.

Reranking

A second-stage model that scores a larger candidate set and keeps only the best few. It improves precision before tokens are spent on prompt context.

Lost in the middle

The tendency for models to attend less reliably to content buried in the middle of a long context. Place the most relevant passages near the beginning or end.

Grounded citations

Tying answer claims to specific retrieved chunk identifiers. Citations make outputs auditable and expose fabricated references that do not map to any source.

Prompt injection

An attack where instructions hidden in untrusted content, such as retrieved documents or user input, attempt to override system rules or hijack tool use.

Defense in depth

Layering input filtering, system-prompt constraints, output validation, and monitoring. Each layer catches failure classes the others miss, so no single control is trusted alone.

Least-privilege tools

Granting an agent only the credentials and permissions its specific task needs. This limits the blast radius if the agent is manipulated or errs.

PII redaction

Removing or masking personal identifiers before content is logged, stored, or sent to third parties. Redacting only at display time leaves data exposed downstream.

Output guardrail classifier

A separate model or rule set that screens generated responses against policy before delivery. It catches violations that prompt instructions alone fail to prevent.

Golden evaluation set

A labeled collection of representative and adversarial cases with expected outcomes. It is the reference for measuring accuracy per category across prompt or model changes.

LLM-as-judge

Using a model to score outputs against a written rubric. It requires calibration against human labels, since uncalibrated judges drift and inflate scores.

End-to-end tracing

Recording prompt version, retrieved chunk IDs, tool calls, token counts, and final output per request. Traces make production failures reproducible and debuggable.

Canary comparison

Routing a small traffic slice to a changed prompt or model and comparing metrics against the baseline. It attributes quality shifts to a specific modification.

Exponential backoff with jitter

A retry strategy that increases wait time between attempts and randomizes it to avoid synchronized retries. Apply only to retryable errors such as rate limits and timeouts.

Model version pinning

Referencing an explicit model version rather than a floating alias. Behavior then changes only on a deliberate, reviewed upgrade rather than silently.