Claude Certified Architect - Professional key terms

The maximum number of tokens a model can attend to in a single request, covering system prompt, conversation history, retrieved documents, and the model's own output. It bounds how much information influences any one response.

A planned allocation of the context window across competing inputs such as instructions, few-shot examples, retrieved passages, and reserved output space. Budgeting prevents any single input from crowding out the others.

Reusing the model's internal computation for a stable prefix of a prompt, such as a long system instruction or reference document, so repeated calls skip reprocessing it. It lowers latency and cost when the prefix stays unchanged.

Stored key and value tensors from prior attention computations that let a model avoid recomputing earlier tokens during generation. Its size grows with sequence length and constrains how many long conversations a server can hold.

Degradation in answer quality as a context fills with marginally relevant material, even when it fits within the window. Models can overlook or mis-weight key facts buried among distractors, so curation matters more than capacity.

An observed pattern where models recall information placed at the start or end of a long context better than material in the middle. Ordering retrieved evidence strategically can improve grounding and answer accuracy.

An instruction pattern that asks the model to reason through intermediate steps before giving a final answer. It improves performance on multi-step problems but increases output tokens and can expose reasoning the user should not see.

Supplying a small number of worked input-output examples inside the prompt to demonstrate the desired format, tone, or reasoning. Examples are strong signals for style and structure but consume context and can bias toward superficial patterns.

The privileged instruction block that sets the model's role, constraints, and output conventions ahead of user turns. It is the primary lever for consistent behavior and for defining refusal and escalation policies.

Constraining a model's response to a machine-readable schema, typically JSON, often enforced by a schema-aware decoding mode. It makes downstream parsing reliable but requires handling of validation failures and truncation.

An attack in which untrusted content, such as a retrieved page or user-supplied document, contains instructions that hijack the model away from its intended task. Defenses include input isolation, least privilege, and output validation.

Prompt injection delivered through data the model ingests rather than through the user's own message, such as a poisoned web page, email, or knowledge-base chunk. It is especially dangerous in agentic systems that can act on the model's output.

Granting a model or agent only the permissions strictly needed for its current task, such as read-only database access or a narrow tool allowlist. It limits the blast radius when the model is manipulated or simply wrong.

A design pattern requiring human approval before an agent takes a consequential or irreversible action, such as sending money or deleting records. It trades throughput for safety on high-stakes steps.

A check that constrains model inputs or outputs, such as content classifiers, schema validators, or policy filters, applied outside the model itself. Guardrails provide deterministic enforcement that prompt instructions alone cannot guarantee.

A deliberate attempt to bypass a model's safety training or policy constraints through role-play, encoding tricks, or adversarial framing. Robust designs layer input filtering, output checks, and monitoring rather than relying on refusal behavior alone.

The capability of a model to emit a structured call to an external function, receive the result, and continue reasoning with it. Tools extend the model beyond its training data and context, enabling actions and fresh information.

The machine-readable description of a tool's name, purpose, and input parameters that the model uses to decide when and how to call it. Precise descriptions and narrow parameter types markedly improve call accuracy.

The data returned from an executed tool call and appended to the conversation so the model can incorporate it. Results must be validated and size-limited, since malformed or oversized outputs can derail subsequent reasoning.

Issuing multiple independent tool calls in a single model turn so they execute concurrently. It reduces round trips and latency, but requires the orchestrator to match results back to calls correctly.

The recurring cycle in which a model observes state, chooses an action or tool call, receives the result, and repeats until a stopping condition is met. Loop design determines cost, latency, and failure modes.

An agent pattern interleaving short reasoning traces with actions and observations, so each step's rationale informs the next tool call. It improves transparency and adaptability at the cost of extra tokens.

A hard cap on how many steps or tool calls an agent may take before the orchestrator halts it and returns partial results or an error. It is a primary defense against runaway loops and unbounded spend.

A control that stops calling a failing dependency, such as a flaky tool or model endpoint, after repeated errors and retries later. It protects latency and cost budgets during partial outages.

Reattempting a failed call after increasing delays, often with jitter, to recover from transient rate limits or errors without overwhelming the service. Only idempotent operations should be retried blindly.

A unique identifier attached to a request so that repeating it produces the same effect rather than duplicating side effects. It is essential when agents retry actions such as payments or record creation.

A topology where a lead model decomposes a task, dispatches subtasks to specialized workers, and assembles their outputs. It scales complex work but adds coordination overhead and partial-failure handling.

A narrower agent invoked by a parent to handle a scoped subtask with its own instructions and tools. Isolating context this way keeps the parent's window clean and limits each worker's permissions.

A dense numeric vector representing text such that semantically similar content lands near each other in vector space. Embeddings power semantic search, clustering, and deduplication but carry no inherent notion of truth.

A measure of the angle between two vectors, used to score how semantically close a query embedding is to a document embedding. Values near one indicate high similarity regardless of vector magnitude.

A store built to index embeddings and answer nearest-neighbor queries at scale, often with metadata filters. It enables semantic retrieval over large corpora but must be kept in sync with source content.

Splitting source documents into smaller passages before embedding and indexing. Chunk size and overlap trade retrieval precision against losing surrounding context that the answer may depend on.

Combining lexical keyword matching with dense semantic retrieval and merging the ranked results. It recovers exact identifiers and rare terms that pure embeddings miss while retaining semantic recall.

A second-stage scoring pass, often with a cross-encoder, that reorders an initial candidate set by relevance to the query. It improves precision before passages enter the context window.

Tying a model's answer to supplied evidence, such as retrieved passages or tool results, and requiring citations or abstention when evidence is absent. It reduces fabrication and makes outputs auditable.

A pattern that retrieves relevant documents at query time and inserts them into the prompt so the model answers from fresh or private knowledge. Quality depends heavily on retrieval, not just the generator.

Confident model output that is unsupported by evidence or factually wrong. Mitigations include grounding, abstention instructions, output validation, and evaluation sets that target fabricated claims.

A curated, versioned collection of inputs with expected outputs or grading criteria used to measure system quality. It must reflect real traffic and be protected from leaking into prompts or training data.

Using a model to score another model's outputs against a rubric, enabling scalable grading of open-ended responses. Judges need calibration against human labels and are vulnerable to position and verbosity bias.

Re-running an evaluation suite after any change to prompts, models, retrieval, or tools to catch quality drops before release. It turns prompt and model changes into measurable, reversible engineering decisions.

Directing each request to an appropriate model tier based on task difficulty, cost, or latency requirements, often with a cheap classifier or fallback chain. It balances quality against spend across a workload.

An ordered sequence of models or providers tried when the primary choice fails, is rate-limited, or returns an invalid response. It improves availability but complicates output consistency and testing.

The maximum acceptable end-to-end response time for a user interaction, decomposed into retrieval, model, and tool time. It drives choices like streaming, caching, smaller models, and parallel calls.

Delivering model output incrementally as tokens are generated rather than waiting for completion. It improves perceived responsiveness but complicates validation, since partial output may later fail schema checks.

Tracking input and output tokens per request, user, and feature to attribute cost and enforce quotas. Accurate accounting requires counting system prompts, retrieved context, tool results, and retries, not just visible text.