Glossary · Plain definitions
The Morvion vocabulary.
Short, plain definitions of the terms we ship against, written for operators who'd rather know what something means than be sold on it. Each entry includes a one-liner, a long-form writeup, and the questions we get asked most often.
AI infrastructure
AI infrastructure is the technical and architectural layer that lets a business run AI-powered workflows reliably in production, retrieval, memory, agents, observability, evaluation, and safety rails.
CRM intelligence
CRM intelligence is the AI-driven layer that enriches, drafts, and acts on data inside a CRM, replacing the passive record system with active agents that score signals, propose next moves, and surface deals before they go quiet.
Immersive website
An immersive website is a digital surface that feels like a coherent space rather than a stack of pages, pacing, motion, materiality, typography, and editorial rhythm tuned together so the visitor moves through an atmosphere instead of a layout.
AI agent
An AI agent is a goal-directed software system that uses an LLM to plan, call tools, and take actions across multiple steps, with measurable outputs that can be scored before they reach a user.
Business intelligence dashboard
A business intelligence dashboard is a real-time decision surface that turns operational data into a small number of visible signals operators can act on, replacing weekly reports with continuous awareness.
Client portal
A client portal is a private digital surface where a service business and its customers run the engagement together, replacing scattered email threads with a single source of truth for projects, files, decisions, and payments.
Discovery sprint
A discovery sprint is a focused two-week engagement that validates the riskiest assumption behind a software, AI, or platform idea with working code and a build-or-don't verdict at the end, both outcomes worth the engagement.
Digital operating layer
A digital operating layer is the connected system that runs a modern business across its surfaces, the visible interface, the data behind it, the AI workflows that move it, and the dashboards that explain it, designed as one stack rather than five disconnected tools.
Document intelligence
Document intelligence is the AI layer that reads, extracts, classifies, and routes the unstructured documents a business runs on, replacing manual triage of contracts, invoices, briefs, and emails with measurable extraction and structured downstream actions.
Eval-driven AI
Eval-driven AI is a development discipline that writes the evaluation harness before the agent, scores every output against a versioned rubric, and ships only on green metrics, the AI equivalent of writing integration tests before a feature.
Hospitality website
A hospitality website is the digital storefront of a venue, bar, restaurant, hotel, or destination, designed to convert atmosphere into reservations, events, and returning guests, with reservations, hours, gallery, and the operational systems wired in.
Marketplace platform
A marketplace platform is a two-sided software system that connects a demand side (requesters, buyers, organisers) with a supply side (operators, sellers, venues) through a matching layer, an operator dashboard, and a trust system designed to move requests across the gap.
Multi-agent workflow
A multi-agent workflow is a production AI system where several role-scoped agents (each with its own tools and rubric) coordinate across a task, replacing a single brittle prompt with a small team of specialists that each do one thing well.
Real-time dashboard
A real-time dashboard is a live data surface that updates within seconds of the underlying event, designed for operators who need continuous awareness of what is changing in the business, not a weekly snapshot.
Retrieval-augmented generation (RAG)
Retrieval-augmented generation is the AI pattern that grounds an LLM's response in a company's own data by retrieving relevant context at query time, replacing model hallucination with citable answers built from the business's knowledge.
Prompt engineering
Prompt engineering is the discipline of structuring instructions to a language model so that outputs are reliable, scoped, and on-task — versioned alongside the rest of the system, not improvised in a chat window.
Vector database
A vector database stores text and other data as high-dimensional embeddings and retrieves the most similar items by mathematical distance — the substrate underneath retrieval-augmented generation, semantic search, and recommendation systems.
AI observability
AI observability is the layer that records, indexes, and replays every prompt, retrieval, tool call, and model response in a production AI system — so operators can debug a single failed output, audit a regulator's question, or measure drift over weeks.
Embedding model
An embedding model converts text, images, or other data into a high-dimensional numerical vector so semantically similar inputs land near each other in space, and a system can ask 'what does this mean' instead of 'does this string match'.
Fine-tuning
Fine-tuning is the practice of taking a pre-trained model and continuing its training on a smaller task-specific dataset, shifting its weights toward the patterns of one domain rather than the general internet.
Vector search
Vector search retrieves the records most semantically similar to a query by comparing their embedding vectors in high-dimensional space, instead of matching keywords.
Semantic search
Semantic search returns results that match the meaning of a query rather than its literal words, using embedding vectors, query rewriting, and reranking under the hood.
Hallucination
An AI hallucination is a confidently-produced statement that has no grounding in the input or in any retrievable source, the failure mode that separates language models from databases and the one production AI must defend against.
Chain-of-thought
Chain-of-thought is a prompting and training pattern that asks an LLM to write its intermediate reasoning steps before its final answer, trading more tokens for higher accuracy on tasks where the steps matter.
Function calling
Function calling is the API pattern where an LLM returns a structured JSON payload requesting that a named function be invoked with specific arguments, so the model can act on external systems instead of only emitting text.
Model distillation
Model distillation is the practice of training a smaller model (the student) to mimic the behavior of a larger model (the teacher) on a specific task, trading a small amount of quality for substantial cost and latency reduction.
Safety rails
Safety rails are the deterministic guards layered around an AI system, input validation, content filtering, refusal handlers, output schema enforcement, rate limiting, so the model fails predictably even when it would otherwise fail strangely.
Eval harness
An eval harness is a deterministic test apparatus for a non-deterministic system: a fixed fixture set, written rubrics, and a scoring run that produces a number comparable across releases.
Regression gate
A regression gate is an automated CI check that compares the current run's eval scores against a saved baseline and fails the PR when any metric drops past its tolerance.
Model Context Protocol (MCP)
The Model Context Protocol is an open standard that lets AI agents discover and call tools, read resources, and use prompts provided by external servers — the interoperable plug for connecting agents to systems they don't natively know.
Structured output
Structured output is the mode in which a language model returns a value matching a declared schema — JSON, a tool call, or a typed object — rather than freeform prose, enforced by the provider's grammar-constrained decoding.
Agent tool use
Agent tool use is the pattern where a language model is given a set of typed functions it can call — search, database queries, code execution, API calls — and the runtime executes the calls and feeds results back, letting the agent operate on real systems.
Prompt injection
Prompt injection is a class of attack where adversarial content in the model's input — direct from a user, or indirect via a retrieved document, image, or webpage — overrides the system prompt and redirects the model to do something its operator did not intend.
Agentic search
Agentic search is the third wave of discovery — AI agents browse the web on a user's behalf, read structured manifests like /mcp-actions.json, and complete tasks directly on sites that expose machine-callable actions, often bypassing the visual UI entirely.
Observability traces
Observability traces are the per-request record of every step an AI system took — model calls, tool invocations, retrieval queries, latencies, token counts — captured automatically and stored for replay, debugging, and regression analysis.
LLM guardrails
LLM guardrails are the deterministic layer around a model that enforces what it cannot do — input validation, content filtering, refusal handling, output schema enforcement, rate limits, tool authorization — so the system as a whole fails predictably.
Agent handoff
An agent handoff is the structured transfer of work from one AI agent to another — a planner to an executor, a router to a specialist — encoded as a typed contract with explicit input, expected output, and success criterion.
Vector index
A vector index is the data structure that makes semantic search fast — it stores embedding vectors with approximate-nearest-neighbour lookups, returning the top-K most similar items in milliseconds across millions of records.
Token budget
The token budget is the declared maximum of tokens an AI workflow may consume per request — for the prompt, the retrieved context, the model's reasoning, and the output combined — enforced at runtime so cost and latency stay predictable.
Retrieval rerank
Retrieval rerank is the second pass over the top-K passages returned by vector search — a small cross-encoder model scores each (query, passage) pair directly and reorders, so the most relevant chunks reach the model first.
Embedding space
Embedding space is the high-dimensional vector geometry where an embedding model places text — semantically similar passages land near each other, and distance is the proxy for relevance.
Semantic cache
A semantic cache stores past prompt-response pairs and serves new requests from the cache when the new request's embedding is sufficiently close to a stored one — letting an AI workflow skip the model call entirely for near-duplicate queries.
Context window
The context window is the maximum number of tokens a language model can read and reason over in a single call — the upper limit on system prompt + conversation history + retrieved context + output combined.
Faithfulness
Faithfulness is the eval metric that measures whether every claim in a model's response is derivable from the retrieved context — the canonical anti-hallucination check for any RAG workflow.
Cross-encoder
A cross-encoder is a model architecture that takes the query and the candidate passage as a single joint input and produces a relevance score — slower than bi-encoder embedding lookup but ten to twenty points more accurate.
Model router
A model router is a small classifier that inspects each incoming query and dispatches it to the right specialist model or workflow — small/fast for easy queries, large/expensive only when the task requires it.
AI cost control
AI cost control is the discipline of budgeting, measuring, and enforcing per-workflow spend on language-model APIs — token budgets, routing, caching, prompt caching, and gateway-level limits combined into a predictable cost ceiling.
Agent memory
Agent memory is the persistent state an AI agent reads at the start of every turn and writes back at the end — the difference between an agent that learns a user across a conversation and one that starts cold every time.
Structured extraction
Structured extraction is the AI workflow that turns unstructured text into a typed object that matches a strict schema — the bridge between free-form documents and the rest of the production stack.
AI evaluation framework
An AI evaluation framework is the discipline-level layer above any single eval harness — the set of fixtures, rubrics, regression policies, and release gates that govern how an AI system is measured across all workflows and over time.
Retrieval quality
Retrieval quality is the family of metrics that measures whether a RAG pipeline actually surfaces the right context for the query — recall@k, precision@k, MRR, nDCG — measured against a labelled fixture set, not vibes.
AI guardrail policy
An AI guardrail policy is the written specification of what an AI system must refuse, must validate, and must escalate — the policy document that the deterministic guardrail code enforces and the eval harness tests against.
Eval fixture
An eval fixture is one input-and-expected-shape pair in an evaluation harness — a real or representative example with the labelled answer or rubric outcome the AI workflow is meant to produce.
Eval rubric
An eval rubric is the written definition of what counts as a good output for one input class — the scoring contract that turns subjective judgement into a number an eval harness can compare across releases.
AI incident
An AI incident is a production failure of an AI system serious enough to warrant a structured response — bad output that reached a user, a regression that blew a release gate, a regulatory exposure, or a runaway cost event.
Agent orchestration
Agent orchestration is the control layer that routes work between named agents, enforces ordering, manages retries, and surfaces the state of the run — the system that turns several role-scoped agents into one production workflow.
Eval versioning
Eval versioning is the discipline of treating the fixture set, the rubric, and the regression baseline as versioned artefacts in git — so every scoreboard is comparable to the one before it, and every drift is traceable to a specific change.
Model fallback
Model fallback is the production pattern of routing to a secondary model when the primary model fails, refuses, or breaches a budget — so a single provider outage or a single transient error doesn't take the workflow down.
Fine-grained routing
Fine-grained routing is the production pattern of dispatching different parts of the same workflow to different models — small/fast for cheap steps, large/expensive only for the steps that need it — so the cost distribution matches the difficulty distribution.
AI policy version control
AI policy version control is the practice of storing the AI guardrail policy in git, reviewing changes through PRs, and pinning every production release to a specific policy version — so every decision the system makes is traceable to the policy it was operating under.