R

RAG (Retrieval-Augmented Generation)

• Category: Retrieval & generation
• What people say: A model answering with retrieved knowledge.
• What it actually means: A system pattern that retrieves evidence relevant to a request and supplies selected content to a generative model before it answers or acts. Retrieval can use lexical, vector, structured, or hybrid methods.
• Why it matters: RAG can make current or private evidence available without encoding it into model weights, but retrieval and grounding must be evaluated separately.
• Why it's called that: Retrieval finds evidence, augmentation adds selected evidence to context, and generation produces the response.
• Learn it: Retrieval-Augmented Generation
• Sources: Retrieval-Augmented Generation paper
• Related terms: Grounding, Hybrid Retrieval, Reranker, Hallucination

Rate Limit

• Category: AI-native development
• What it actually means: A policy that caps requests, tokens, concurrent work, or another resource within a defined time or capacity window.
• Why it matters: It protects providers and your own system from overload, uncontrolled spend, and unfair resource use.
• In practice: Enforce per-tenant token and concurrency limits, read provider retry metadata, and queue or reject excess work predictably.
• Common confusion: A rate limit controls allowed usage. Backpressure propagates downstream capacity constraints through a system.
• Related terms: Backpressure, Retry with Backoff, Circuit Breaker

ReAct

• Category: Agents & tools
• What it actually means: An agent pattern that interleaves task reasoning, a concrete action, and an observation returned by the environment before deciding the next step.
• Why it matters: Environment feedback can correct assumptions and ground later decisions instead of forcing the model to complete the entire task from internal generation alone.
• In practice: Expose a small set of typed tools, return concise observations, cap the loop, and verify the final artifact rather than storing private reasoning traces.
• Common confusion: ReAct is a prompting and control pattern, not a guarantee of autonomy, correctness, or safe tool use.
• Related terms: Agent, Function Calling, Planning, Grounding
• Sources: ReAct

Readiness Probe

• Category: Reliability & operations
• What it actually means: A diagnostic that tells the traffic-routing layer whether a service instance is currently able to accept requests.
• Why it matters: A process can be alive while its model is unloaded, dependencies are unavailable, or warmup is incomplete, so sending traffic too early creates avoidable failures.
• In practice: Check the minimum dependencies required to serve, fail readiness during startup and draining, keep the probe inexpensive, and do not restart the process solely because readiness is false.
• Common confusion: Readiness controls traffic eligibility. Liveness decides whether the process should be restarted, and neither proves that every model response will be correct.
• Learn it: Production LLM Application
• Related terms: Autoscaling, Model Serving, Availability, Graceful Degradation
• Sources: Kubernetes Liveness, Readiness, and Startup Probes

Recall@K

• Category: Retrieval & generation
• What it actually means: For one query, Recall@K is |relevant items intersecting the top k| / |relevant items|. A dataset score aggregates those per-query values under a stated rule.
• Why it matters: It tells you whether a retrieval stage supplies downstream generation or reranking with enough relevant candidates.
• In practice: Define relevance judgments, k, the aggregation method, and a policy for queries with no judged relevant items, then inspect queries with zero recalled evidence.
• Common confusion: High Recall@K does not mean the top result is good, the ranking is well ordered, or the final answer is grounded. Queries with no relevant items require an explicit exclusion or assigned-value policy because the denominator is zero.
• Related terms: Precision & Recall, Eval Set, Reranker, Approximate Nearest Neighbor (ANN)
• Sources: BEIR

Reciprocal Rank Fusion (RRF)

• Category: Retrieval & generation
• What it actually means: A rank-fusion method that combines several result lists by summing contributions that decrease with each item's rank in each list.
• Why it matters: It can merge lexical, dense, or multi-query rankings without assuming their raw scores share the same scale.
• In practice: Retrieve independent candidate lists, deduplicate by stable document identity, apply one versioned fusion constant, and evaluate against each individual retriever.
• Common confusion: RRF combines ranks, not embeddings or relevance scores, and it cannot recover an item absent from every input list.
• Related terms: Hybrid Retrieval, BM25, Dense Retrieval, Reranker
• Sources: Reciprocal Rank Fusion Outperforms Condorcet and Individual Rank Learning Methods

Red Teaming

• Category: Security & governance
• What it actually means: A structured adversarial testing process in which authorized testers seek failures using documented objectives, threat assumptions, cases, and evidence.
• Why it matters: Ordinary quality tests rarely explore how a system behaves under manipulation, misuse, conflicting goals, or determined attempts to bypass controls.
• In practice: Derive attacks from a threat model, run them in an isolated environment, record reproducible cases, remediate by layer, and convert confirmed failures into regression evals.
• Common confusion: A list of jailbreak prompts is not a complete red-team program, and red teaming cannot prove the absence of unknown failures.
• Related terms: Threat Model, Guardrails, Prompt Injection, Eval Set
• Sources: Red Teaming Language Models with Language Models

Regression Test

• Category: AI-native development
• What it actually means: A repeatable check that protects behavior known to work, especially after code, prompt, model, retrieval, or tool changes.
• Why it matters: AI system changes can improve average quality while silently reintroducing a previously fixed failure.
• In practice: Turn a corrected prompt-injection incident into a permanent eval case that must pass before the next deployment.
• Common confusion: A regression test guards a specific expected behavior. A broad benchmark estimates performance across a wider task distribution.
• Learn it: Eval-Driven Agent Development
• Related terms: Eval Set, Verification Gate, Patch, Evaluation (Eval)

ReLU

• Category: Math & training
• What people say: A simple activation function.
• What it actually means: Rectified Linear Unit, defined as f(x) = max(0, x). It is inexpensive and has a non-saturating positive branch, though zero gradients on negative inputs can create inactive units.
• Related terms: Activation Function, Gradient, CNN (Convolutional Neural Network)

Repository Instructions

• Category: AI-native development
• What it actually means: Version-controlled guidance that tells coding agents how a repository is organized, which commands and conventions apply, what boundaries to respect, and how to verify work.
• Why it matters: It turns repeated tribal knowledge into local context that travels with the code and can vary by subproject.
• In practice: Keep an AGENTS.md at the repository root, add narrower files for subdirectories, and include exact build, test, generated-file, security, and contribution rules.
• Common confusion: Repository instructions complement source code and human documentation; they do not override the user's current request or guarantee that an agent follows them correctly.
• Related terms: Repository Map, Scope Contract, Coding Agent, Progressive Disclosure
• Sources: AGENTS.md specification

Repository Map

• Category: AI-native development
• What it actually means: A compact, maintained description of a repository's important directories, ownership boundaries, entry points, build commands, tests, generated files, and local instructions.
• Why it matters: It helps a coding agent find the right evidence before loading large files or editing the wrong subsystem.
• In practice: Generate an index from the tree and manifests, then enrich it with authoritative notes about module boundaries and validation commands.
• Common confusion: A raw file tree shows names. A repository map explains which paths matter and how they relate to a task.
• Learn it: Repository Memory and State
• Related terms: Coding Agent, Progressive Disclosure, Scope Contract, Context Engineering

Reproducible Build

• Category: AI-native development
• What it actually means: A build whose declared source, environment, and instructions can be independently rerun to produce bit-for-bit identical specified artifacts.
• Why it matters: It makes an artifact verifiable beyond the machine or agent that originally produced it and exposes hidden build inputs.
• In practice: Pin toolchains and dependencies, remove timestamps and unstable ordering, capture the environment, then compare independently rebuilt artifact digests.
• Common confusion: A build that succeeds twice is repeatable evidence, but reproducibility requires the declared independent conditions and identical outputs.
• Related terms: Repository Instructions, Verification Gate, Provenance Attestation, Software Bill of Materials (SBOM)
• Sources: Reproducible Builds definition

Reranker

• Category: Retrieval & generation
• What it actually means: A second-stage model or scoring function that reorders a small candidate set using a richer comparison between the query and each candidate.
• Why it matters: Fast first-stage retrieval maximizes candidate coverage, while reranking can improve which evidence reaches the limited context window.
• In practice: Retrieve 50 candidates with hybrid search, score each query-document pair with a cross-encoder, and pass the top 5 supported chunks to generation.
• Common confusion: A reranker does not search the entire corpus. It only reorders candidates that retrieval already found.
• Related terms: Hybrid Retrieval, Semantic Search, RAG (Retrieval-Augmented Generation)

Retry Budget

• Category: Reliability & operations
• What it actually means: A bound on retry traffic, usually expressed relative to original requests or over a time window, that prevents retries from consuming unbounded capacity.
• Why it matters: When a dependency slows or fails, unrestricted retries multiply load exactly when the system has the least spare capacity.
• In practice: Count retries separately from first attempts, cap them by service and tenant, honor deadlines, use jittered backoff, and stop retrying non-transient or non-idempotent failures.
• Common confusion: A retry budget limits extra attempts. An error budget measures user-visible unreliability allowed by an SLO.
• Related terms: Retry with Backoff, Error Budget, Rate Limit, Admission Control
• Sources: Google SRE: Addressing Cascading Failures

Retry with Backoff

• Category: AI-native development
• What it actually means: Repeating a failed transient operation after progressively longer delays, usually with randomized jitter and a strict retry limit.
• Why it matters: Immediate synchronized retries can worsen an outage, consume rate limits, and duplicate side effects.
• In practice: Retry a provider timeout after bounded exponential delays, honor server retry guidance, and reuse an idempotency key for any write.
• Common confusion: Do not retry permanent validation or permission errors, and do not retry non-idempotent operations without a duplication strategy.
• Related terms: Idempotency, Rate Limit, Circuit Breaker, Backpressure

Reviewer Agent

• Category: AI-native development
• What it actually means: An agent assigned to inspect another agent's artifact or decision against explicit criteria and return findings or a verdict.
• Why it matters: Separation of roles can catch omissions, but it only helps when the reviewer receives independent evidence and a concrete rubric.
• In practice: After one agent produces a patch, give a separate reviewer the diff, scope contract, repository rules, and test output, then require line-specific findings.
• Common confusion: A second model call is not automatically independent or correct. Shared context, model bias, and vague criteria can reproduce the same mistake.
• Learn it: Reviewer Agent
• Related terms: Coding Agent, Verification Gate, Scope Contract, LLM-as-a-Judge

RLHF (Reinforcement Learning from Human Feedback)

• Category: Math & training
• What people say: Training a model from human preferences.
• What it actually means: A family of pipelines that uses human feedback to learn a reward or preference signal and then optimizes a model policy against that signal. Implementations vary and need not all use the same reinforcement-learning algorithm.
• Common confusion: RLHF optimizes a proxy learned from collected feedback. It does not guarantee broad alignment with every user or situation.
• Learn it: Reinforcement Learning from Human Feedback
• Sources: InstructGPT paper
• Related terms: DPO (Direct Preference Optimization), SFT (Supervised Fine-Tuning), Alignment

Rollback

• Category: Reliability & operations
• What it actually means: Restoring a previously known deployment or configuration when the current release violates operational, quality, or safety criteria.
• Why it matters: Agent and model changes can fail in production despite pre-deployment evaluation, so recovery must be designed before rollout.
• In practice: Retain versioned artifacts and configuration, define rollback triggers, rehearse the command and data implications, and verify service health after restoration.
• Common confusion: Code rollback does not automatically reverse database migrations, external side effects, cached outputs, or data written by the bad release.
• Related terms: Canary Release, Checkpoint, Regression Test, Durable Execution
• Sources: Kubernetes Deployments: Rolling Back

ROUGE

• Category: Evaluation & safety
• What people say: A reference-overlap metric often used for summaries.
• What it actually means: A family of metrics that compares generated text with reference text using units such as n-gram overlap or longest common subsequence.
• Common confusion: Surface overlap can miss semantic equivalence and can reward copied wording without proving factual quality.
• Related terms: Evaluation (Eval), Precision & Recall, LLM-as-a-Judge