NVS Dictionary

The Automation Dictionary.

Plain-language definitions for the terms we use across this site. When we mention something that isn't everyday language, we link it here so you can check what it means without leaving the page you're on.

/
AI & Retrieval
Agentic RAG agentic-rag A version of RAG where the AI decides for itself when to search, how to rephrase the question, and whether to search again, rather than retrieving once and stopping. Benchmark benchmark A standard test set and scoring method used to compare AI or retrieval systems on the same task, so claims can be measured instead of asserted. BM25 bm25 A long-standing keyword search method that ranks documents by how well their exact words match the query. Simple, fast, and still hard to beat on precise text. Chunk chunk A small passage a document is split into so it can be searched and retrieved on its own. Chunking is the practice of doing that splitting well. Context window context-window The fixed amount of text a language model can hold in mind at once for a single request. Its working memory, measured in tokens. Embeddings embeddings A way of turning text into a list of numbers that captures its meaning, so a computer can measure how closely two pieces of text are related. F1 score f1-score A single accuracy number that balances precision and recall into one figure, used to compare systems on a task. Higher is better. Groundedness groundedness How much of an AI's answer is actually backed by the documents it retrieved. Measured claim by claim, it is the honest test of whether an answer can be trusted. Hallucination hallucination When an AI states something that sounds confident and fluent but isn't supported by any real source. The main failure mode that RAG and evaluation are built to catch. Hybrid retrieval hybrid-retrieval Searching with keyword matching and meaning-based search at the same time, then merging the two result lists so each covers the other's blind spots. LLM llm Large language model. The kind of AI, trained on huge amounts of text, that reads a prompt and writes a fluent response. The engine underneath most modern AI tools. Prompting prompting Shaping the input you give a language model so it produces the output you want. The everyday craft of getting a useful answer out of an AI. RAG rag A way of answering questions with an AI where the system first looks up relevant documents, then writes its answer from what it found instead of from memory alone. Recall@k recall A retrieval score: how often the correct document showed up in the search's top k results. Recall@5 means 'was the right passage in the top 5?', measured across many questions. Reranker reranker A second-pass model that re-reads the top search results against the question and re-orders them by true relevance, pushing the best passage to the top. Sampling sampling How a language model picks its next word: by drawing from a range of likely options rather than always taking the single most likely one. It is also where invented details creep in. Semantic search semantic-search Search that matches on meaning instead of exact words, so it can find the right passage even when it shares no keywords with your question. Token token The small chunk of text (roughly a word-piece) that a language model reads and generates. It is also the unit AI usage is measured and billed in. Vector store vector-store A database built to store embeddings and quickly find the ones most similar to a given piece of text.
Information Theory
Bandwidth bandwidth How much a channel can carry, its width. In Shannon's theory it sets the ceiling on information rate; here it stands for any fixed capacity you cannot cheaply enlarge, like a model's context window. Bit bit The fundamental unit of information: one yes-or-no choice. Entropy, and the cost of resolving uncertainty, are counted in bits. Channel capacity channel-capacity The hard ceiling on how much information a channel can carry reliably. From Shannon's work, and the lens behind how we think about scaling a business's throughput. Entropy entropy A measure of uncertainty. In information theory, the amount of surprise in an outcome, counted in bits, and the reason some tasks cost more to pin down than others. Entropy floor entropy-floor The irreducible minimum uncertainty in a task: the point past which no amount of prompting or retrieval can remove ambiguity, so it has to be resolved by supplying more information. Information theory information-theory The branch of math, founded by Claude Shannon, that measures information and the limits of moving it accurately through any channel. The framework beneath how NVS thinks about automation. Mutual information mutual-information How much knowing one thing reduces your uncertainty about another. The precise way to say how many bits a retrieved document adds to an answer. Redundancy redundancy The predictable, repeated part of a message, the portion above its entropy floor. It is what makes text compressible and errors correctable. Signal-to-noise signal-to-noise The ratio of useful information to background clutter. Raising it, by removing noise rather than adding volume, is how a channel carries more. Source coding source-coding Compressing a message down to its essential information and no further. Shannon proved you can shrink a source to its entropy, but not below it without losing something.
Automation & Systems
API api A defined way for two pieces of software to talk to each other automatically, sending data and triggering actions without a person clicking through a screen. API key api-key A secret string that identifies and authorizes an app when it calls an API. Like a password for software, so it has to be stored carefully and never exposed. Audit log audit-log A tamper-evident, append-only record of every action a system took, who did what and when, kept for accountability and for tracing a problem back to its cause. CRM crm Customer relationship management software: the system of record for a business's clients, leads, and deal history that automations read from and write back to. Durable execution durable-execution A workflow-engine guarantee that a long-running process resumes exactly where it left off after a crash, restart, or long wait, instead of starting over or losing its place. Enrichment enrichment Automatically adding known details to a record, like filling in a lead's company, role, and size from just their email address. Error handling error-handling The built-in logic for what a system does when a step fails: catch it, report it, retry or pause, rather than breaking silently or stopping the whole job. Guardrails guardrails Enforced limits and approval checks placed around an automation so it can only act within safe, allowed bounds, no matter what it is asked to do. Health check health-check An automatic test that regularly confirms a system is running and reachable, so a failure is caught and alerted before customers notice. Human-in-the-loop human-in-the-loop A design where an automation pauses for a person's approval on the decisions that carry real risk, while handling the safe, routine ones on its own. Idempotency idempotency A property that makes it safe to run the same step twice: doing it again has no extra effect, so a retry can't accidentally double-charge or double-send. JSON json A common text format for organizing data as labeled fields and values, structured so software can read and pass it around reliably. OAuth oauth A standard that lets you grant an app limited access to your account on another service without handing over your password. It hands out a revocable key instead. Observability observability The ability to see what a running system is actually doing from the outside, through logs, traces, and metrics on every run, so problems are visible instead of silent. Orchestration orchestration The layer that runs workflows: deciding what runs when, retrying steps that fail, and keeping long or multi-step jobs coordinated and reliable. Pipeline pipeline A chain of automated steps where each stage's output feeds the next, running end to end with no manual handoff in between. Relational database relational-database A system for storing data in linked tables, so related records like customers and their orders connect to each other and can be queried exactly. Retry retry Automatically re-attempting a step that failed the first time, usually after a short wait, so a temporary glitch does not become a dropped task. Scraping scraping Automatically pulling information off web pages so a program can use it, instead of copying it by hand. Self-healing self-healing A system that detects its own failures and automatically retries or pauses the affected step to recover, without a person stepping in for the routine cases. Shadow mode shadow-mode Running a new automation for real but with its outward actions held back, so you can watch exactly what it would have done before letting it touch a customer. Smoke test smoke-test A quick automated check that runs a system end to end with fake data to confirm the whole path still works, before real traffic hits it. Source of truth source-of-truth The single system designated as the authoritative record for a piece of data, so every tool and person reads the same value instead of keeping separate copies that drift apart. Stack stack The full set of software tools and services a business runs on, taken together as one connected setup. When people say 'your stack,' they mean all your apps as a whole. Tenant tenant In shared software, one customer's isolated slice of the system, walled off so one customer's data is never visible to another. Trigger trigger The event that starts an automation running, like a new email arriving, a form being submitted, a scheduled time, or a record changing. Webhook webhook An automatic message one system sends another the moment something happens, so a workflow can start instantly instead of checking over and over for changes. Workflow workflow A defined sequence of automated steps that carries a task from trigger to finish, like turning an order intake into a confirmation, a record, and a calendar event.
No term matches that.