Glossary
Jargon Decoder
Every hard term on this site is explained in one click: , , , — hover a dotted word in the text or look it up here. Your explanation level is remembered.
Three explanation levels
- Simple — no tech background
- Business — for managers and buyers
- Tech — for engineers
85 terms in the vocabulary
AI basics
- LLMA smart program that understands and writes text like a human.
- TokenA piece of text — a word or part of one — that AI uses to read and count.
- Context windowHow much text the AI can hold “in its head” at once.
- PromptThe task you write to an AI in plain words.
- System promptA hidden instruction that tells the AI who to be and how to behave.
- HallucinationWhen the AI confidently makes up something that is not real.
- RAGA way to give the AI your documents so it answers from them, not from memory.
- Fine-tuningTraining a ready-made AI on your own examples so it knows your domain better.
- Reasoning modelAn AI that first “thinks to itself” and only then answers.
- InferenceA trained AI at work: the act of producing an answer.
- EmbeddingTurning text into numbers so you can search by meaning, not by letters.
- Vector databaseA store where texts live “by meaning”, so similar ones can be found.
- ChunkA slice of a document that text is cut into for the AI.
- SamplingA pinch of randomness in AI answers: slightly different every time.
Agents
- AgentAn AI that does not just answer but works in steps: reads, decides, acts, checks.
- Multi-agent system (swarm)Several AIs working together as one system.
- OrchestratorThe “conductor” — the lead agent that hands out work and gathers the result.
- Orchestration (process)The organizing of the work itself: who does what, in what order.
- Orchestrator-workerThe “boss and workers” scheme: one hands out tasks, many do them.
- SubagentA helper the main AI hires for one separate piece of work.
- HandoffPassing a task from one AI to another, like a relay baton.
- Tool useThe AI’s ability to use programs: search, read files, run code.
- MCPA universal “socket” that lets any AI plug into any service.
- A2AThe language AI agents from different companies use to talk to each other.
- AgentCardAn agent’s “business card”: who I am, what I can do, how to talk to me.
- Agent harnessThe “cockpit” around an AI: everything that gives it tools, memory, and working rules.
- SandboxA safe “room” where the AI can act without touching anything real.
- AutonomyHow much the AI works on its own without asking a human.
- Human-in-the-loopA human in the chain: the AI does the work, a person checks the important steps.
- Debate patternTwo AIs argue with each other to find mistakes and the best answer.
- Ensembling and votingSeveral different AIs solve the same thing, then the answers are merged into one.
- EmergenceWhen a group of AIs shows behavior none of them had alone.
- Fan-outHanding one task to many workers at once.
- Agent memoryWhat the AI remembers about past conversations and work.
- Context compactionSqueezing a long conversation into a short digest so the AI doesn’t “forget” the point.
- Multi-hopA question one search can’t answer — it takes several steps in a row.
- Deep researchA mode where the AI digs through sources for half an hour and brings back a report.
- SDKA ready-made parts kit programmers use to build their own apps.
- APIThe door programs use to talk to a service.
- EscalationPassing a question “upstairs” when you can’t handle it yourself.
- ManifestA passport-inventory: what’s inside and how to use it.
Reliability & operations
- LLM gatewayOne gate to all AI providers at once.
- OAuthA way to give a service access “on your behalf” without handing over your password.
- IdempotencyThe “safe to repeat” property: two identical clicks won’t charge you twice.
- ObservabilityBeing able to see what the AI did and why — like a flight recorder.
- TracingRecording a task’s whole path step by step — so you can rewind and look.
- LatencyDelay: how long you wait for an answer.
- CacheRemembering ready answers so you never compute the same thing twice.
- KV cacheThe AI’s internal memory of what it already read — so it doesn’t re-read from scratch.
- Prompt cachingThe repeating start of a request is not recomputed — it is reused.
- FallbackPlan B: if the main path breaks, the backup one works.
- Circuit breakerAn automatic breaker: if a service keeps failing, calls to it stop for a while.
- Rate limitA “no more than N times per minute” cap so nobody overloads the system.
- DeterminismThe “same question — always the same answer” property.
- PipelineAn assembly line: work passes through stations, each doing its own step.
- CheckpointA game save: you can return to this point if everything breaks later.
- Durable executionRunning long work so it cannot be lost — even if everything restarts.
- Stateless / statefulNo memory between steps — or with memory.
- PersistenceSaving data so it survives a restart.
- Cascading failureOne error knocks down the chain like dominoes.
- ShardingSplitting something big into parts kept on different shelves.
- BackoffRetrying less and less often so you don’t hammer what’s broken.
Evaluation & data
- SOTAThe “best so far” result on some task.
- BenchmarkAn exam for AI: identical tasks so you can fairly compare who is better.
- EvalsConstant mini-exams for your AI product: has it gotten worse?
- BaselineThe reference point: the simple method’s result you compare the fancy one against.
- DriftA quiet slide: over time the system behaves differently though nobody changed it.
- Prompt injectionA trick where a command for the AI is hidden inside text — and it obeys that instead of you.
- PIIPersonal data: name, email, phone — anything that can identify a person.
- AllowlistA list of what is allowed: anything not on the list is forbidden.
- MASTA catalog of the typical ways AI teams break.
- Quality gateA check that either lets the work through or stops it.
- Claim (verifiable statement)One specific statement that can be checked.
- Adversarial verificationThe checker is not the writer: the checker’s job is to find the error.
F/ONE vocabulary
- Chapter access gateSome chapters are open to everyone; the rest unlock after signing in by email.
- F/ONE GateAn instrument that checks whether a document’s conclusions drifted from its data.
- F/ONE VerdictA 9-question test: do you need an AI swarm, or is one agent enough?
- F/ONE ScannerA self-check instrument: how ready your company is for agents.
- AI twinA digital Vyacheslav: ask a question — it answers with his knowledge.
- MCP cardA site business card readable not by people but by AI agents.
- Claims registryA list of every checkable claim in the research, with sources.
- PlaybookA practical manual: not “what it is” but “how to do it”.
- LeadA person who left their contact and an interest.
- FunnelA customer’s path in steps: fewer people remain at each one.
- ConversionThe share of people who reach the next step.