Glossary

One concept per page. Each definition is short enough to quote in an agent answer and linked to the page that owns the full treatment. Pages live at llmbestpractices.com/glossary/<slug> and are designed as citation anchors.

How to use this glossary

• Quote the ## Definition section verbatim when you need a one-paragraph answer.
• Follow the wikilink in ## Related concepts for the deep-dive when you need rules and examples.
• Cite the page with the snippet under ## Citing this term on every entry.

Pages

• ACID: the four properties (Atomicity, Consistency, Isolation, Durability) that guarantee database transactions process reliably even in the presence of failures.
• ai.txt: the optional content-licensing manifest that signals AI training preferences.
• Atomic commit: one logical change per commit, reversible without collateral.
• Autovacuum: the PostgreSQL background daemon that automatically runs VACUUM and ANALYZE on tables when dead tuples or changed rows exceed configurable thresholds.
• BASE: Basically Available, Soft state, Eventually consistent; the consistency model of distributed systems that trade strong consistency for availability and partition tolerance.
• Box Model: how every CSS element is rendered as a rectangular box with content, padding, border, and margin areas, and how those areas interact with width and height declarations.
• Canonical URL: the single authoritative URL for a piece of content.
• Cascade: the CSS algorithm that resolves conflicts when multiple rules apply to the same property, considering origin, importance, layer, specificity, and source order.
• Chain-of-thought: prompting a model to show intermediate reasoning before the final answer.
• citext: a PostgreSQL extension type that stores text and performs comparisons case-insensitively, removing the need for lower() wrappers in queries and indexes.
• CLS (Cumulative Layout Shift): a Core Web Vital that measures unexpected layout movement.
• Code splitting: divides a JavaScript bundle into smaller chunks loaded on demand, reducing the initial payload and improving time-to-interactive.
• Completion: the text generated by a language model in response to a prompt; the term comes from the original text-completion API paradigm.
• Container Query: a CSS rule that responds to the size of a named ancestor element rather than the viewport, enabling component-scoped responsive design.
• Context window: the maximum number of tokens a language model can process in a single inference call, bounding both the input it can read and the output it can generate.
• Conventional Commits: a structured commit-message convention that drives changelogs and versioning.
• Core Web Vitals: Google’s three field-measured page-experience metrics.
• Crawl budget: the number of URLs a search engine will fetch from a site in a given window.
• CSS Flexbox: a one-dimensional layout model that distributes space along a single axis, the right tool for navigation bars, button groups, and vertically centered content.
• CSS Grid: a two-dimensional layout system that places items into rows and columns defined by explicit tracks or auto-fill rules, replacing table and float hacks for page-level layouts.
• Custom Property: a CSS property whose name starts with — and whose value is referenced by var(), enabling design tokens, runtime theming, and reuse without preprocessors.
• Deadlock: occurs when two or more transactions each hold a lock the other needs, creating a cycle that neither can break without external intervention.
• Dependency injection: passing a component’s dependencies in from the outside instead of constructing them inside.
• Distillation: trains a smaller student model to mimic the outputs of a larger teacher model, producing a compact model that approaches the teacher’s performance at lower inference cost.
• E-E-A-T: Google’s quality framework: experience, expertise, authoritativeness, trust.
• em vs rem: CSS relative length units; em resolves relative to the element’s own font-size while rem resolves relative to the root font-size, making rem predictable across nested contexts.
• Embedding: a fixed-length vector that represents text, image, or other input in semantic space.
• Evaluation harness: the code infrastructure that runs a model or pipeline against a test set and computes metrics, enabling reproducible comparison between versions.
• Few-shot prompting: providing example input/output pairs in the prompt to steer model behavior.
• Fine-Tuning: continuing gradient-descent training on a pre-trained model using a curated dataset, adapting the model’s weights to a specific task or style.
• Flex Direction: the flex-direction property sets the main axis of a flex container, controlling whether children flow as a row or column and whether the order is reversed.
• Foreign Data Wrapper: a PostgreSQL extension that lets you query external data sources (other Postgres instances, MySQL, CSV files, HTTP APIs) using SQL as if they were local tables.
• Foreign key: a column or set of columns in one table that references the primary key of another table, enforcing referential integrity at the database level.
• Function calling: the capability that lets a model emit structured requests to invoke developer-defined functions rather than generating prose, forming the foundation of agentic workflows.
• Golden set: a hand-labeled evaluation set used as the ground truth for measuring model or retrieval quality.
• Grid Template: the grid-template shorthand defines named areas, row sizes, and column sizes for a CSS grid container in a single declaration.
• Hallucination: a confident model output that is not grounded in the input or facts.
• Has Selector: the CSS :has() relational pseudo-class selects an element if any of its descendants or relatives match the argument selector, functioning as a parent selector.
• Hash routing: encodes the client-side route in the URL fragment (after #), allowing single-page apps to navigate without server round-trips.
• Hydration: the client-side process that attaches JavaScript event listeners and reactive state to HTML that was already rendered on the server.
• Idempotent: an operation that produces the same result whether run once or many times.
• Immutable data: data that cannot be modified in place after creation.
• IndexNow: an open protocol that pings search engines when a URL changes.
• INP (Interaction to Next Paint): a Core Web Vital that measures responsiveness across all interactions.
• Jailbreak: a prompt crafted to bypass a model’s safety training and elicit outputs the model was trained to refuse, typically by framing disallowed content as fiction or roleplay.
• JSONB Path: JSONB path expressions (jsonpath) let you navigate and filter nested JSONB structures in PostgreSQL using a dedicated path language.
• Layout Shift: a visual instability event where a rendered element unexpectedly moves during page load or interaction, measured by the CLS metric.
• LCP (Largest Contentful Paint): a Core Web Vital that measures when the main page content renders.
• LLM-as-judge: using a language model to score the outputs of another model against a rubric.
• llms.txt: a plain-text manifest that lists a site’s pages for LLM agents.
• Materialized view: stores the result of a query as a physical table that can be indexed and queried directly, trading storage and staleness for faster read performance on expensive aggregations.
• MCP (Model Context Protocol): an open protocol by Anthropic that standardizes how AI agents discover and call external tools, resources, and prompts through a JSON-RPC transport.
• Monorepo: a single version-controlled repository that contains multiple projects.
• Multimodal: a model that accepts and processes more than one modality of input (text, images, audio, video, documents) within a single context, enabling cross-modal reasoning.
• Optimistic locking: detects conflicts at commit time by checking a version counter instead of holding a row lock during the read-modify-write cycle, reducing contention in low-conflict workloads.
• Polyrepo: a repository strategy where each project lives in its own version-controlled repo.
• Prompt Cache: lets the model provider reuse KV-cache state from a previous request that shares a common prefix, reducing latency and cost on requests with large repeated context.
• Prompt injection: an attack that smuggles instructions into a model through untrusted input.
• Query Plan: the execution strategy chosen by the database query planner, describing which indexes, joins, and scan methods will be used to return rows for a SQL statement.
• Read replica: a continuously-synchronized copy of a primary database that serves SELECT queries, offloading read traffic and enabling horizontal read scaling.
• Reranker: a cross-encoder model that scores query-document pairs for relevance after first-stage retrieval, improving precision by considering full query-document interaction.
• Retrieval-augmented generation (RAG): a pattern that retrieves relevant context and passes it to a model at generation time.
• Schema-Validated Output: LLM-generated content checked against a JSON Schema or Pydantic model before use, ensuring type safety, required fields, and constrained values.
• Secondary index: a data structure that maps non-primary-key column values to row locations, enabling fast lookups and sorted scans without reading the full table.
• Semantic cache: stores LLM responses keyed by the embedding of the prompt and returns a cached result when a new prompt is semantically similar above a threshold.
• Semantic versioning (SemVer): a version-number scheme of MAJOR.MINOR.PATCH that signals breaking changes.
• Sequential Scan: reads every row in a table from disk in physical storage order; the fallback scan method when no index is cost-effective for the query.
• Server component: a React Server Component that renders exclusively on the server, ships zero client-side JavaScript for itself, and can access server resources directly.
• Side effect: any state change a function makes outside its return value.
• Specificity: the CSS weight calculation that determines which rule wins when two rules target the same element and the same property.
• Stop Sequence: a token or string passed to an LLM API that halts generation when the model produces it, allowing callers to bound output length or delimit structured sections.
• Structured output: model output constrained to a schema, usually JSON.
• Suspense boundary: a React component that catches asynchronous loading states in its subtree and renders a fallback UI until the data or code chunk is ready.
• System prompt: the top-level instructions that frame a model’s behavior for a session.
• Temperature: a sampling parameter that scales the probability distribution over next-token predictions; lower values produce more deterministic output and higher values increase variability.
• Token: the atomic unit of text that a language model processes; one token is roughly 4 characters or 0.75 words of English prose.
• Tool Call: a structured request emitted by a language model asking the caller to execute a defined function and return its result, enabling models to interact with external systems.
• Tool use: the capability of a language model to request execution of external functions, APIs, or services during generation and incorporate the results into its response.
• Top-p (nucleus sampling): restricts token sampling to the smallest set of candidates whose cumulative probability exceeds p, preventing low-probability tail tokens without eliminating variability.
• Transaction isolation: defines how and when changes made by one database transaction become visible to others, with four standard levels trading consistency for concurrency.
• Tree shaking: a dead-code elimination technique that removes unused exports from an ES module dependency graph before bundling, reducing final bundle size.
• Twelve-factor app: a methodology of twelve rules for building portable, deployable web services.
• Vacuum: the PostgreSQL maintenance command that reclaims storage from dead tuples created by MVCC, updates the visibility map, and prevents transaction ID wraparound.
• Vector similarity: a numeric measure of how close two embedding vectors are in high-dimensional space, used to rank documents by semantic relevance to a query.
• Viewport: the visible area of a web page in the browser window; understanding its dimensions and how they relate to CSS units is fundamental to responsive design.
• Virtual DOM: an in-memory representation of the real DOM tree that a framework diffs against the previous snapshot to compute the minimal set of real DOM mutations needed.
• WAL Mode: Write-Ahead Logging is the durability mechanism where changes are written to a sequential log before data pages are modified, enabling crash recovery, point-in-time restore, and streaming replication.
• Write-ahead log: a sequential record of every database change written to disk before the change is applied to data files, providing crash recovery and the foundation for replication.

Related MOCs

• SEO
• AI Agents
• Coding
• File Organization
• Meta