AB-620 flashcards

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

Deterministic logic placement

Rule stating that fixed, rule-based calculations and validations belong in application code rather than in a model prompt, because code is testable and produces identical output every time.

Latency budget

The maximum acceptable end-to-end response time, decomposed across model inference, retrieval, and network overhead so each component has a measurable allocation.

Context window

The total number of tokens a model can process in one request, covering system prompt, retrieved content, conversation history, and the generated output combined.

RAG versus fine-tuning

Decision rule: retrieval suits changing or citable facts, while fine-tuning suits stable knowledge where the gap is tone, format, or task-specific behavior.

Chunk overlap

A repeated portion of adjacent text chunks, typically ten to twenty percent, that prevents answers from being lost when relevant content straddles a chunk boundary.

Human-in-the-loop trigger

A condition that routes a model output to a person for approval, such as low confidence, an irreversible action, or a high-risk intent classification.

Tool schema

The typed definition of a function exposed to an agent, including parameter names, types, descriptions, and required flags that the model uses to form valid calls.

Agent iteration cap

An explicit maximum on plan-act-observe cycles in an agent loop, preventing runaway token spend and infinite retry behavior when tools keep failing.

Workflow versus agent

Design choice: use a fixed workflow when steps are known in advance, and an autonomous agent only when the next step depends on runtime results.

Grounded response

An answer whose factual claims are supported by a retrieved record or tool response, with the supporting source identified rather than asserted from model memory.

Idempotency key

A caller-supplied identifier on a state-changing tool call that lets the service detect retries and avoid creating duplicate records or repeating side effects.

Structured handoff summary

A compact object passed between agents containing intent, constraints, and collected facts, replacing the raw transcript to control token growth and preserve focus.

Fallback chain

An ordered sequence of alternatives, such as primary model then secondary model then scripted response, each hop logging its reason for later analysis.

Retrieval pipeline order

The fixed sequence of ingest, chunk, embed, index, query embedding, similarity search, optional rerank, and prompt assembly used in retrieval-augmented generation.

Hybrid search

Combining vector similarity with keyword matching and merging ranked results, which improves recall for queries containing exact identifiers, codes, or product names.

Embedding model consistency

The requirement that query and document vectors come from the same model and version, meaning the corpus must be re-embedded whenever the embedding model changes.

Exponential backoff with jitter

A retry strategy that increases wait time between attempts and adds randomization, reducing synchronized retry storms when a service is throttling requests.

Chunk metadata

Fields attached to each chunk such as source id, title, section path, timestamp, and access tags, enabling citation, freshness checks, and per-user filtering.

Event-driven index update

Indexing triggered by source-system change events rather than a nightly batch, which shortens the staleness window between source updates and retrievable content.

Data contract validation

Checking incoming payloads against a defined schema, nullability, and value ranges before sending them to a model, rejecting payloads that fail validation.

Groundedness metric

An evaluation score measuring the proportion of claims in a response that are supported by the provided context or retrieved sources.

Retrieval recall

The fraction of queries whose relevant chunk appears within the top-k retrieved results, a leading indicator of downstream answer quality.

Regression gate

A release rule requiring that a prompt or model change not push any per-category evaluation metric below its defined minimum threshold.

Drift detection

Monitoring input topic and output length distributions against a baseline, triggering re-evaluation when observed shifts exceed expected variation.

Optimization order

A prioritization sequence for quality work: retrieval quality first, then prompt clarity, then model selection, then caching, because retrieval errors dominate failures.

Prompt caching TTL

A time-to-live on cached model responses tied to data freshness, disabled for personalized or permission-filtered answers that must not be reused across users.

Single-variable A/B test

An experiment changing exactly one factor, such as prompt, model, temperature, or retrieval depth, so any measured difference can be attributed to that factor.

Prompt injection defense

Treating retrieved documents and user input as untrusted data rather than instructions, with tool permissions enforced outside the model in code.

Least privilege agent identity

A design in which each agent holds only the scopes its tools require, so a compromised or manipulated model cannot reach data beyond the workflow's needs.

PII redaction

Detecting and removing sensitive personal fields before sending content to a model, and logging identifiers or hashes instead of raw values.

Canary release

Deploying a new model or prompt version to a small slice of live traffic first, monitoring quality and error signals before full promotion.

Audit logging

Recording prompt version, model version, tool calls, and decision outcome for each request so any output can be reconstructed and reviewed later.

Layered content safety

Applying input filtering, model-side safety settings, and output filtering together, since no single control reliably blocks all harmful content.