Claude Certified Architect - Professional flashcards

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

Stateless orchestration

Architecture where conversation and task state live in an external store rather than process memory, allowing any instance to serve any request and enabling horizontal scaling.

Idempotency key

A unique identifier derived from the logical request that lets a tool or API detect and discard duplicate retries, preventing repeated side effects such as double charges.

Control plane vs data plane

The control plane holds configuration, prompts, routing, and policy; the data plane performs inference and retrieval. Separating them lets prompt changes ship without redeploying serving infrastructure.

Graceful degradation

Predefined fallback behavior when a model times out, is rate limited, or retrieval misses, such as serving a cached answer or returning an explicit uncertainty response.

Context budget

The allocation of the model's context window across system prompt, retrieved chunks, tool schemas, conversation history, and reserved output tokens before retrieval depth is chosen.

Model tiering

Routing requests to different model sizes based on task difficulty, sending classification and extraction to small fast models and multi-step reasoning to frontier models.

Prompt caching

Reusing computed representations of a stable prompt prefix to cut cost and latency; invariant content goes first, variable content last to maximize cache hits.

Structured output

Constraining model responses to a JSON schema or strict tool schema so machine consumers get parseable results, with validation and repair on parse failure.

Few-shot prompting

Including a small set of input-output examples in the prompt to establish format and edge-case handling, typically three to five diverse examples rather than many near-duplicates.

Context summarization

Compressing older conversation turns into a summary to free context space while preserving key facts, an alternative to sliding windows or full history replay.

Lost-in-the-middle effect

The tendency for models to attend more reliably to information at the beginning and end of long contexts, so critical instructions should be placed after bulk content.

Tool schema

A typed JSON definition giving a tool's name, description, and parameters; precise descriptions are the main lever for correct tool selection by the model.

Agent loop

The cycle of model emitting a tool call, runtime validating and executing it, returning the result as a tool message, and the model continuing, bounded by an iteration cap.

Orchestrator-worker pattern

A lead model decomposes a task into subtasks, dispatches them to worker calls possibly in parallel, then synthesizes results into a final answer.

Evaluator-optimizer pattern

One model generates a draft while another scores it against criteria and returns feedback, iterating until a quality threshold is met; used for quality-critical outputs.

Human-in-the-loop gate

A required human approval step before an irreversible or high-blast-radius action such as a payment, deletion, or external message is executed by an agent.

Chunk overlap

Duplicated text between adjacent retrieval chunks, typically 10-20 percent, that prevents information loss when a relevant passage straddles a chunk boundary.

Hybrid retrieval

Combining dense vector similarity with lexical search such as BM25, which improves recall for exact identifiers, product codes, and rare terms that embeddings handle poorly.

Reranking

Rescoring a larger candidate set from initial retrieval with a cross-encoder to improve precision before the top few chunks are placed in context.

Recall@k

The fraction of queries for which at least one relevant document appears among the top k retrieved results, measuring retrieval coverage independently of answer quality.

Metadata pre-filtering

Applying tenant, date, source, or permission constraints inside the vector search rather than after it, preserving recall and enforcing access control.

Golden dataset

A curated set of real queries with labeled expected outputs, typically 50 to 200 cases, used as the reference for offline evaluation and regression testing.

LLM-as-judge

Using a model with a defined rubric to score outputs at scale, requiring calibration against human labels and a fixed judge version for comparable scores over time.

Faithfulness check

Verifying that each claim in a generated answer is supported by the retrieved context, catching hallucination that surface accuracy metrics can miss.

Retrieval vs generation failure

Diagnostic split: a wrong answer with correct context indicates a generation failure, while a wrong answer with missing context indicates a retrieval failure.

Regression suite

A test set that grows with every production incident and previously failing case, run before each change to prevent reintroducing known defects.

Prompt injection

An attack where untrusted content such as a retrieved document or user input contains instructions the model follows, exploiting a failure to separate data from instructions.

Least-privilege tools

Narrowly scoped tools with credentials limited to exactly the operations needed, so a compromised or misdirected agent cannot perform broad actions.

Server-side authorization

Enforcing access checks at tool execution and retrieval time using the caller's verified identity, rather than trusting claims made inside the prompt.

Audit logging

Recording model version, prompt template version, retrieved sources, and tool calls for each consequential decision so outputs can be reviewed and traced.

Time to first token

The latency until the first streamed output token arrives; reducing it improves perceived responsiveness even when total generation time is unchanged.

Semantic cache

A cache that returns a stored response when a new query's embedding is sufficiently similar to a prior one, cutting cost for near-duplicate requests.