AI & LLMs

Busted

• AgentA model in a loop: it picks a tool, reads the result, decides the next move, and keeps going until the goal is met or the budget runs out.
• Agentic workflow; agent orchestrationAn AI assembly line that routes a job through model calls, tools, branches, and review steps instead of asking one model once.
• Batch inferenceRunning many independent model requests as a queued job, trading instant answers for lower cost or higher throughput.
• Chain-of-thought promptingPrompting a model to work through intermediate steps before giving its answer, especially for problems where each step depends on the last.
• ChunkingCutting documents into retrievable pieces before embedding them. Too big and hits go vague; too small and the answer loses the thread.
• Context overflow; truncationWhen a request is too large for the model's token window. Overflow is the failure; truncation is your rule for dropping content to fit.
• Context rotThe model getting less reliable as its context fills with stale, noisy, or competing material, even though the request still fits.
• Context windowThe maximum tokens a model can process at once, prompt plus reply. Not a memory: a desk. Overflow is an error, not a polite trim.
• Cosine similarityA score for how closely two embedding vectors point in the same direction, regardless of how long the vectors are.
• DistillationTraining a small, cheap model to imitate a larger teacher, trading some general ability for lower cost and faster responses.
• EmbeddingA list of numbers that encodes what text means, so similar meanings land close together and you can search by sense instead of spelling.
• EvalsUnit tests for prompts: a fixed set of inputs, scored the same way every run, so you can prove a change made things better and not worse.
• Few-shot promptingPutting two or three worked examples in the prompt so the model copies the pattern instead of guessing what you meant.
• Fine-tuningTraining an existing model further on your own examples so the behaviour is baked in and you stop repeating it in every prompt.
• Foundation modelA broadly pretrained model used as the starting point for many products, then adapted with prompts, retrieval, or fine-tuning.
• Golden datasetThe trusted answer key for your AI tests: curated real inputs paired with the exact result or behavior the system should produce.
• GroundingForcing the answer to come from sources you supplied, with citations pointing at them, not from what the model vaguely recalls.
• GuardrailsChecks wrapped around a model that block bad input going in and bad output coming out, enforced in code rather than politely requested in the prompt.
• HallucinationOutput nothing actually supports: invented facts, fake citations, a function that does not exist. Confident or hedged, it's still made up.
• Human in the loopThe agent pauses and asks a person to approve before doing anything real: sending, spending, deleting, deploying.
• Hybrid searchRunning keyword search and vector search at once and merging the hits, so you catch exact terms and reworded ideas alike.
• InferenceActually running a trained model to produce an answer. Training builds the model; inference is every live use of it afterward.
• Input tokensAll tokens the model receives before replying, including hidden instructions, chat history, documents, examples, and tool definitions.
• JailbreakA prompt designed to talk a model past its learned safety rules and make it produce something it would normally refuse.
• Knowledge cutoffThe rough date where a model's built-in training knowledge stops being current. Newer facts need live sources, not confident guessing.
• LLM-as-judgeUsing one model as the grader for another model's answers when quality is too fuzzy for an exact test.
• LoRA adapterA small trained add-on that changes a frozen base model's behaviour without saving or retraining a whole new copy of the model.
• Max tokensThe hard ceiling on how many tokens the model may write back. It limits the response, but does not ask the model to use them all.
• MCP (Model Context Protocol)An open standard for plugging tools and data into AI apps, so one server works with any assistant instead of a bespoke integration each.
• MemoryFacts saved outside the chat and reloaded into later ones, so the model appears to remember you after the conversation resets.
• Metadata filteringNarrowing retrieval by facts like tenant, date, language, or document type before ranking the remaining candidates by meaning.
• Model routerA traffic controller that sends each AI request to the cheapest model capable of doing that particular job.
• Model snapshotA pinned, versioned model release that will not silently move when a provider updates its floating model alias.
• MultimodalA model that takes more than text: images, audio, video, PDFs. You paste the screenshot instead of describing it.
• Open-weights modelA model whose learned parameter files you can download and run yourself. The license, training code, and training data may still be closed.
• Output tokensThe pieces a model generates for its reply. They count toward cost and latency, and an output limit can cut the answer short.
• PromptThe instructions and input you hand a model for one turn. Anything specific to your situation exists only if you put it in there.
• Prompt cachingGetting the provider to remember the unchanging front of your prompt, so you stop paying full price and full latency for the same instructions.
• Prompt injectionText the model reads (a doc, a page, an email) that gives it new orders and it obeys, because instructions and data look identical to it.
• Prompt templateA reusable prompt with named blanks that code fills at runtime, so one tested instruction can handle many inputs consistently.
• QuantizationStoring model weights with fewer bits so they use less memory and can run faster locally, at the cost of some numerical precision.
• Query rewritingTurning a messy or conversational request into better search queries before retrieval, while preserving what the user actually meant.
• Reasoning modelA model that burns extra computation working a problem out before it replies. Slower and pricier, much better on the hard stuff.
• RerankingA second, slower model re-scores your top search results and reorders them, so the genuinely best chunk lands first.
• Semantic cachingReusing a previous answer when a new request means nearly the same thing, even though the words are different.
• Semantic searchSearch that matches on meaning instead of exact words, so 'how do I cancel' finds the page titled 'ending your subscription'.
• Stop sequenceA literal piece of text that tells the generation engine to stop as soon as the model produces it.
• StreamingSending the answer out token by token as it is generated, so text types itself onto the screen instead of landing all at once.
• Structured outputGetting data back instead of prose: a JSON object that matches a schema you defined, so your code can parse it without regex archaeology.
• SubagentA worker agent sent off with one scoped task and its own context, then asked to return a result to the parent agent.
• Synthetic dataTraining or test examples made by a model instead of collected from people or production, useful for scale but only after hard filtering.
• System promptYour app's own instructions, sent ahead of the user's and weighted above them: role, rules, tone. Usually repeated every turn.
• TemperatureThe randomness dial on word choice. Low picks the safest next token every time; high lets unlikely ones through, for better or worse.
• Time to first token (TTFT)How long you stare at nothing before the first character appears. The wait before the model starts typing, not how long it takes to finish.
• TokenThe chunk of text a model actually reads, usually a word piece, not a whole word. Limits, speed and price are all counted in these.
• Token budgetThe token allowance you choose for a request or session. It caps cost and size below the model's absolute context-window limit.
• TokenizerThe model-specific converter that splits text into numbered token pieces and turns generated token IDs back into text.
• Tool calling / function callingThe model can't run anything itself. It emits a request naming a function plus arguments, your code runs it, and the result goes back in.
• Top-p / nucleus samplingA word-choice cutoff: keep only the smallest group of likely next tokens whose probabilities add up to p, then sample from that group.
• Vector databaseA store built to answer 'what is closest in meaning to this?' It indexes embeddings so nearest-neighbor search stays fast at millions of rows.
• Zero-shot promptingAsking a model to do the task without showing it any worked examples first. The prompt is instructions and input only.

The territory

30 core terms mapped for this field, ranked by how often builders reach for them. Each one is a future entry. Want to bust one? One entry, one file, one pull request.

• PromptInput or instruction sent to a model for a response"what I type in" · "the question I ask it"
• System promptPersistent instruction block that sets a model's role and rules"the hidden instructions" · "the personality setup text"
• Context windowMax tokens a model can hold in working memory at once"how much it can remember" · "the size limit of the conversation"
• RAG (retrieval-augmented generation)Fetch relevant documents, inject them into the prompt before answering"make it read my docs" · "let it search my files first"
• EmbeddingNumeric vector representing meaning, enabling similarity search"turning text into numbers" · "the meaning fingerprint"
• Vector databaseStore searching by semantic similarity rather than keywords"search by meaning not words" · "the AI memory store"
• ChunkingSplitting documents into retrievable pieces before embedding"cutting docs into bits" · "how big the snippets are"
• Tool calling / function callingModel emits structured requests your code executes and returns"let it press buttons" · "make it use my API"
• Structured outputForcing responses into a strict JSON or schema shape"make it always return JSON" · "stop it rambling, give me fields"
• AgentLLM looping over tools and observations toward a goal"AI that does stuff on its own" · "it keeps going till done"
• MemoryPersisted facts recalled across separate sessions"it remembers me" · "saves what I told it last time"
• HallucinationConfident, fluent output that is factually invented"it made that up" · "it lies convincingly"
• TemperatureSampling knob trading determinism for creativity and variety"how random it gets" · "the creativity dial"
• TokenSub-word unit models read, generate, and bill by"chunks of words" · "the thing they charge for"
• Few-shot promptingIncluding worked examples so the model mimics the pattern"show it examples first" · "teach by demonstration in the prompt"
• Fine-tuningFurther training a base model on your own examples"train it on my data" · "make it learn my style permanently"
• Prompt injectionUntrusted input that hijacks the model's instructions"someone talks it into ignoring rules" · "malicious text in the doc"
• Reasoning modelModel that spends extra hidden compute thinking before replying"the slow smart one" · "the one that thinks first"
• MCP (Model Context Protocol)Standard for connecting models to tools and data sources"plug my app into Claude" · "the USB port for AI tools"
• StreamingEmitting tokens progressively so text appears as it generates"the typewriter effect" · "words appearing live"
• Time to first token (TTFT)Delay before the first character appears"how long before it starts typing" · "the wait before anything shows up"
• GuardrailsChecks constraining what a model may accept or output"stop it saying bad things" · "the safety fence"
• EvalsTest suite scoring model outputs against expected behavior"unit tests for prompts" · "how do I know it got better"
• Prompt cachingReusing a repeated prompt prefix to cut cost and latency"stop paying for the same instructions" · "remember the setup part"
• Semantic searchRetrieving by meaning similarity instead of exact keywords"search that understands intent" · "finds it even with different words"
• Hybrid searchCombining keyword and vector retrieval for better recall"keyword plus meaning search" · "use both kinds of search"
• RerankingSecond-pass model reordering retrieved chunks by relevance"sort the search hits again" · "put the best snippet first"
• MultimodalModel accepting images, audio, or video alongside text"it can see pictures" · "upload a screenshot and ask"
• GroundingTying answers to supplied sources with citations"make it cite where it got that" · "answer only from these docs"
• Human in the loopRequiring person approval before an agent acts"ask me before it does anything" · "approval step"

Deeper in the field

• Zero-shot prompting Asking directly with no examples provided
• Top-p / nucleus sampling Probability-mass cutoff controlling word choice diversity
• Stop sequence String that halts generation when produced
• Max tokens Hard cap on response length per request
• Context overflow; truncation Dropping content once the window is exceeded
• Model router Dispatching each request to the cheapest capable model
• Quantization Shrinking model weights for cheaper, faster local inference
• LoRA adapter Small trained layer stacked on a base model
• Inference Actually running a model to produce output
• Jailbreak Prompt that bypasses a model's safety training
• LLM-as-judge Using a model to grade another model's outputs
• Golden dataset Curated input-output pairs used as the eval benchmark
• Agentic workflow; agent orchestration Multi-step pipeline coordinating models, tools, branches
• Subagent Delegated agent handling a scoped task with its own context
• Knowledge cutoff Date beyond which the model knows nothing
• Synthetic data Model-generated training or test examples
• Chain-of-thought prompting Making the model reason step by step before answering
• Distillation Training a small cheap model to mimic a large one
• Context rot Quality degrading as the context window fills with noise
• Open-weights model Model whose parameters you can download and self-host
• Token budget Allowance of tokens you spend per request or session
• Prompt template Reusable prompt structure populated with variables at runtime
• Tokenizer Component that converts text between strings and model tokens
• Input tokens Tokens supplied to the model as context
• Output tokens Tokens generated by the model in its response
• Foundation model General-purpose pretrained model adapted for many downstream tasks
• Model snapshot Fixed, versioned release of a model with stable behavior
• Metadata filtering Restricting retrieval using document attributes before similarity ranking
• Query rewriting Reformulating a request to improve retrieval quality
• Cosine similarity Common measure of directional closeness between embedding vectors
• Semantic caching Reusing prior answers for meaningfully similar requests
• Batch inference Processing many model requests together for lower cost or higher throughput