The working
studio glossary.

Retrieval-Augmented Generation: an LLM technique where the model retrieves relevant documents before generating a response.

The process of training a base LLM further on your own data to specialize its outputs.

A vector representation of text, image, or other data used for similarity search.

A database optimized for storing and querying high-dimensional vectors (embeddings).

A search technique combining keyword (BM25) and semantic (vector) retrieval.

A second-stage model that reorders retrieved results by relevance to the query.

A test suite for AI features that measures quality, regressions, and edge cases.

An LLM-powered system that autonomously chooses tools and takes multi-step actions.

An LLM's ability to call external functions (search, calculator, database, etc.) during a response.

A structured way for LLMs to invoke developer-defined functions with typed arguments.

LLM responses constrained to a JSON schema or specific format.

Using one LLM to evaluate the outputs of another LLM (or itself) against criteria.

The practice of crafting LLM inputs to produce better, more reliable outputs.

A prompting technique where the model is asked to reason step by step before answering.

API-level caching of prompt prefixes to reduce cost and latency on repeated calls.

The maximum number of tokens an LLM can process in a single call.

The unit of text an LLM processes — typically 3-4 characters or about 0.75 of a word.

A parameter (0-2) controlling how random or deterministic an LLM's output is.

A sampling parameter that limits LLM output to the smallest set of tokens whose probabilities sum to p.

Low-Rank Adaptation: a lightweight fine-tuning method that trains small adapter layers on top of a frozen base model.

Fine-tuning a model on labeled input-output pairs.

Reinforcement Learning from Human Feedback: training a model based on human preference rankings of outputs.

A simpler alternative to RLHF that trains directly on preference pairs without a reward model.

Quantized LoRA: combines LoRA with 4-bit quantization to fine-tune large models on consumer GPUs.

Reducing the numerical precision of model weights to make inference cheaper and faster.

A compact LLM (typically 1-15B parameters) optimized for specific tasks or constrained environments.

An architecture where each forward pass routes tokens through a subset of "expert" sub-networks.

The neural network architecture underlying virtually every modern LLM.

The transformer component that lets each token in a sequence attend to other tokens.

The initial training of a foundation model on massive amounts of unlabeled text.

A large pretrained model that can be adapted to many downstream tasks.

The most capable AI models — currently models like Claude Opus, GPT-4, and Gemini Ultra.

An LLM trained or post-trained to perform multi-step reasoning, often using extended hidden thinking tokens.

An LLM that can process more than text — images, audio, video, or structured inputs.

Logging, tracing, and monitoring of LLM calls in production.

Adversarial testing of an AI system to find harmful, biased, or wrong outputs.

A prompt that bypasses an LLM's safety training to make it produce restricted content.

An attack where malicious instructions in user input or retrieved data hijack the LLM's behavior.

When an LLM generates plausible-sounding but factually incorrect information.

Standardized evaluation suites for comparing AI models on common tasks.

An open protocol for connecting LLMs to external tools, data sources, and contexts.

Protocols and patterns for AI agents to discover, communicate, and coordinate with each other.

A system where multiple AI agents collaborate or compete on a task.

The process of running an already-trained model to produce predictions or generations.

A runtime cache of attention key/value tensors that speeds up sequential token generation.

An inference technique using a smaller "draft" model to propose tokens that a larger model verifies.

A quantized model file format used for efficient CPU and GPU inference, popularized by llama.cpp.

Running AI models on devices (phones, browsers, IoT) rather than in the cloud.

A degradation in LLM quality as context length grows, even within the model's stated window.

Runtime checks that validate or filter LLM inputs and outputs against policies.

The full cost of running an AI feature in production — inference, eval, observability, ops, and people.

Generating large numbers of SEO-optimized pages from templates and structured data.

Specific, lower-volume search queries that collectively make up most search demand.

Identifying search queries to target based on volume, difficulty, intent, and business fit.

The underlying goal behind a user's search query — informational, navigational, commercial, transactional.

Experience, Expertise, Authoritativeness, Trustworthiness — Google's framework for content quality.

Your Money or Your Life — content where accuracy can affect health, finances, or wellbeing.

A Google ranking system that demotes pages written primarily for SEO rather than humans.

Pages with little or no unique value — typically targeted in pSEO penalties.

Search Engine Results Page — what users see after a search.

Google's AI-generated summaries that appear at the top of SERPs for many queries.

A proposed standard file telling LLM crawlers how to find and use a site's content.

Optimizing content for AI-powered search interfaces (ChatGPT, Perplexity, Claude, AI Overviews).

A boxed answer at the top of a SERP, pulled from one ranking page.

Structured data added to a page to help search engines understand its content.

A JSON-based format for embedding structured data (like schema markup) in web pages.

Google's page experience metrics: LCP, INP (formerly FID), and CLS.

A file listing all important URLs on a site, submitted to search engines.

The number of URLs a search engine will crawl on a site within a given timeframe.

The process of search engines storing a page in their database for retrieval in search results.

A link tag indicating the preferred URL when multiple URLs serve the same content.

The HTML title element — the primary signal for what a page is about and the headline shown in SERPs.

A 150-160 character summary of a page that often appears under the title in SERPs.

Optimizations applied directly to a page — content, headings, links, schema, meta tags.

Links between pages on the same site — a major signal of topical authority and crawl depth.

A link from another site pointing to yours — a major off-page ranking signal.

A Moz/Ahrefs-style aggregate score (0-100) estimating a domain's ability to rank.

The hierarchical structure of a site — how URLs and internal links are organized.

The pattern of URL paths — affects SEO, UX, and shareability.

A comprehensive top-level page on a topic, supporting and linked to by many narrower cluster pages.

A pillar page plus its supporting cluster pages — a coordinated set on one topic.

A content architecture pairing one comprehensive pillar page with many supporting cluster pages on a topic.

A coordinated set of design components, tokens, patterns, and guidelines used across a product.

Named, semantic values for design properties — colors, spacing, typography — shared across platforms.

The complete set of brand assets, rules, and applications — beyond just a logo.

The visual elements — logo, type, color, imagery — that represent a brand.

A coordinated set of logo variants — primary, secondary, monogram, wordmark, lockups — for different uses.

A coordinated set of typefaces, weights, and rules for typography across a brand.

Rules for how UI elements animate — duration, easing, choreography — applied consistently across a product.

A tool for building and documenting UI components in isolation.

Rendering web pages on the server before sending HTML to the client.

Pre-rendering web pages to HTML at build time, served as static files.

A hybrid pattern where static pages are regenerated on-demand or on a timer.

Running render or API logic in CDN-edge locations close to the user, not in a single origin server.

Content Delivery Network — globally distributed servers that cache and serve content close to users.

Core Web Vital measuring how fast the largest visible element loads — should be under 2.5s.

Core Web Vital (since March 2024) measuring responsiveness — should be under 200ms.

Core Web Vital measuring unexpected layout shifts during page load — should be under 0.1.

A React framework for production-grade web apps — by Vercel.

A JavaScript library for building component-based UIs — by Meta.

A typed superset of JavaScript that compiles to plain JavaScript.

A utility-first CSS framework that scales by composing small classes.

An architecture pattern: pre-rendered JavaScript, APIs, and Markup served from a CDN.

A single repository housing multiple related projects — apps, packages, services.

A runtime toggle for enabling or disabling features without re-deploying.

Comparing two variants of a UI or experience by randomly assigning users to each.

When a product satisfies a market deeply enough that demand pulls it forward.

Tracking groups of users (cohorts) over time to measure retention, behavior, and lifetime value.

The ratio of customer lifetime value to customer acquisition cost — a core SaaS health metric.

A framework focusing on the underlying job a customer is trying to get done, rather than demographics.