# AI engineering glossary URL: https://momentaj.com/glossary/ One-sentence definitions of the AI engineering terms Momentaj uses, each with the pages where it comes up. ## AI agent URL: https://momentaj.com/glossary/#ai-agent An AI agent is a language model that can call tools to act on real systems, such as searching, reading a record or placing an order, on behalf of a person. Most of an agent's quality is in its tools, not the model: a few of them, named and described so the model picks the right one, each returning compact results. Anything irreversible should wait for a person's approval. Related: - [AI agents and multi-agent systems](https://momentaj.com/services/ai-agents/) - [Multi-agent systems in .NET: what we learned](https://momentaj.com/insights/multi-agent-systems-in-dotnet-what-we-learned/) - [Notes on AI agents](https://momentaj.com/insights/tag/ai-agents/) ## Multi-agent system URL: https://momentaj.com/glossary/#multi-agent-system A multi-agent system is one in which several AI agents, each with its own role and tools, work on the same task. Add an agent when the work has a genuinely different mandate, not because the task is big. DevGuardian AI runs three, an Architect, a QA reviewer and a Security reviewer, because each owns a different question about the same change. Related: - [Multi-agent systems in .NET: what we learned](https://momentaj.com/insights/multi-agent-systems-in-dotnet-what-we-learned/) - [DevGuardian AI case study](https://momentaj.com/case-studies/devguardian-ai/) - [Notes on AI agents](https://momentaj.com/insights/tag/ai-agents/) ## Tool calling URL: https://momentaj.com/glossary/#tool-calling Tool calling is how a language model acts on real systems: it picks one of the tools it has been given, such as search, read a record or place an order, supplies the arguments, and reads the result before it answers. Tool results should be small: in MayAI a small service sat between the live restaurant data and the agent to reduce each response to plain records, because every field sent to the model costs time and money. We trace every tool call from the first prototype and set a budget for its cost and latency. Related: - [AI agents and multi-agent systems](https://momentaj.com/services/ai-agents/) - [Keeping AI agents honest: the server-side grounding pattern](https://momentaj.com/insights/keeping-ai-agents-honest-server-side-grounding/) - [Notes on AI agents](https://momentaj.com/insights/tag/ai-agents/) ## Reply schema (structured output) URL: https://momentaj.com/glossary/#reply-schema A reply schema is a fixed structure, usually a JSON object, that an AI agent must fill in for every reply instead of free text, so the server can build what the user sees from its parts. In MayAI every reply holds the agent's reasoning, a message for the user, up to five quick-reply options and the identifiers of the items to show, but not the items themselves. With server-side grounding behind it, a reply schema is a constraint rather than a request, which is why it holds where prompts do not. Related: - [Keeping AI agents honest: the server-side grounding pattern](https://momentaj.com/insights/keeping-ai-agents-honest-server-side-grounding/) - [MayAI case study](https://momentaj.com/case-studies/mayai/) - [AI agents and multi-agent systems](https://momentaj.com/services/ai-agents/) ## Approval gate URL: https://momentaj.com/glossary/#approval-gate An approval gate is a point in an AI system at which anything irreversible waits for a person to approve it before the agent acts. The agent proposes and a person approves, and the system makes approval easy rather than possible to skip, with every action, tool call and reasoning step logged. On an MCP server, actions come only once agents have used the read tools and an approval step exists. Related: - [AI agents and multi-agent systems](https://momentaj.com/services/ai-agents/) - [AI readiness for Canadian businesses: a checklist](https://momentaj.com/insights/ai-readiness-checklist-for-canadian-businesses/) - [Should your product have an MCP server?](https://momentaj.com/insights/should-your-product-have-an-mcp-server/) ## Model Context Protocol (MCP) URL: https://momentaj.com/glossary/#mcp The Model Context Protocol (MCP) is an open standard that lets AI hosts such as ChatGPT, Claude, VS Code and Gemini CLI connect an agent to a product's tools. A product takes part through an MCP server. The specification changed often in its first two years; its 2026-07-28 revision is a stateless redesign, and each host adds rules of its own on top. Related: - [MCP servers and AI integrations](https://momentaj.com/services/mcp-development/) - [Your website has a new visitor: AI agents](https://momentaj.com/insights/your-website-has-a-new-visitor/) - [Notes on MCP servers](https://momentaj.com/insights/tag/mcp/) ## MCP server URL: https://momentaj.com/glossary/#mcp-server An MCP server is the program that lists a product's tools, such as search, read a record or create an order, so an AI agent can call them on a person's behalf. A good first server has three to six well-described tools, reads before it writes, returns compact results, and uses OAuth 2.1 so the agent can never do more than the person it acts for. Related: - [Should your product have an MCP server?](https://momentaj.com/insights/should-your-product-have-an-mcp-server/) - [Testing an MCP server](https://momentaj.com/insights/testing-an-mcp-server-what-exists-and-what-is-missing/) - [Notes on MCP servers](https://momentaj.com/insights/tag/mcp/) ## AI host URL: https://momentaj.com/glossary/#ai-host An AI host is the application that runs MCP servers for a person, such as Claude, ChatGPT, VS Code, Cursor or Gemini CLI. Each host adds its own rules to the MCP specification, on tool-name length, annotations, timeouts and result sizes, and the rules conflict. A server meant for several hosts is designed to the strictest of them. Related: - [What each AI host requires of an MCP server](https://momentaj.com/insights/what-each-ai-host-requires-of-an-mcp-server/) - [Notes on MCP servers](https://momentaj.com/insights/tag/mcp/) ## Tool annotations URL: https://momentaj.com/glossary/#tool-annotations Tool annotations are the hints an MCP server attaches to each of its tools, such as read-only, destructive and open-world, that tell an AI host how the tool behaves. They are optional in the MCP specification, but hosts enforce them. As of 1 October 2026, Claude's directory requires a title and a read-only or destructive hint on every tool, and ChatGPT requires all three hints, set explicitly on every tool, and names hints that do not match a tool's behaviour as a common reason for rejection. Whether a read-only hint is honest takes judgment, not a static check. Related: - [What each AI host requires of an MCP server](https://momentaj.com/insights/what-each-ai-host-requires-of-an-mcp-server/) - [MCP servers and AI integrations](https://momentaj.com/services/mcp-development/) - [Notes on MCP servers](https://momentaj.com/insights/tag/mcp/) ## Retrieval-augmented generation (RAG) URL: https://momentaj.com/glossary/#rag Retrieval-augmented generation (RAG) answers questions with a language model by first retrieving the relevant pieces of your own material. Most RAG systems split documents by token count and rely on vector similarity, which fails on structured material such as code, contracts and catalogues. Parsing the material into its real units and retrieving with more than one signal works better. Related: - [RAG and knowledge systems](https://momentaj.com/services/rag-knowledge-systems/) - [RAG for code and other structured material](https://momentaj.com/insights/rag-for-code-and-structured-material/) - [Notes on retrieval (RAG)](https://momentaj.com/insights/tag/rag/) ## Chunking URL: https://momentaj.com/glossary/#chunking Chunking is splitting material into the pieces a retrieval system indexes and returns; done well, each chunk is a unit of meaning, such as a clause, a section, a table, a class or a method, rather than a fixed number of tokens. Most RAG systems split documents by token count, and a chunk of 500 tokens cut from a source file starts in the middle of one method and ends in the middle of another. In DevGuardian AI, source is parsed with Roslyn so that each chunk is a class, a method, an interface or a SQL statement, with metadata that says what it is and where it lives. Related: - [RAG and knowledge systems](https://momentaj.com/services/rag-knowledge-systems/) - [RAG for code and other structured material](https://momentaj.com/insights/rag-for-code-and-structured-material/) - [Notes on retrieval (RAG)](https://momentaj.com/insights/tag/rag/) ## Hybrid search URL: https://momentaj.com/glossary/#hybrid-search Hybrid search combines more than one retrieval signal for each query, such as semantic similarity, lexical or symbolic matching and graph traversal, so that vector similarity is one signal among several. In DevGuardian AI, semantic similarity, symbolic matching on names and signatures, and graph traversal are combined for each query. Lexical search can come from Lucene, Azure AI Search or PostgreSQL full text. Related: - [RAG and knowledge systems](https://momentaj.com/services/rag-knowledge-systems/) - [RAG for code and other structured material](https://momentaj.com/insights/rag-for-code-and-structured-material/) - [DevGuardian AI case study](https://momentaj.com/case-studies/devguardian-ai/) ## GraphRAG URL: https://momentaj.com/glossary/#graphrag GraphRAG is retrieval-augmented generation that also follows the relationships between pieces of material, stored in a graph, rather than relying on text similarity alone. In DevGuardian AI the codebase's dependencies are modelled in Neo4j, so an agent reviewing a change can ask what else breaks, not only what looks similar. Related: - [RAG for code and other structured material](https://momentaj.com/insights/rag-for-code-and-structured-material/) - [DevGuardian AI case study](https://momentaj.com/case-studies/devguardian-ai/) - [Notes on retrieval (RAG)](https://momentaj.com/insights/tag/rag/) ## Server-side grounding URL: https://momentaj.com/glossary/#server-side-grounding Server-side grounding is a pattern in which an AI agent returns only identifiers and the server builds everything the user sees from real tool results, so the agent cannot show an invented item. We built it for MayAI, where a dish that is not on the menu or a price that is not real would be an order nobody can fulfil. It is a constraint rather than a prompt, so it holds when the model changes, and it can be tested. Related: - [Keeping AI agents honest: the server-side grounding pattern](https://momentaj.com/insights/keeping-ai-agents-honest-server-side-grounding/) - [MayAI case study](https://momentaj.com/case-studies/mayai/) ## Guardrails URL: https://momentaj.com/glossary/#guardrails Guardrails are the controls that limit what an AI system can do and show, such as approval gates, grounding, defences against injection, an audit log, and cost and latency budgets. Setting them starts from what a wrong answer costs, in money, in a customer or in a regulator's attention. The data and guardrail requirements are a deliverable of the first stage of an engagement, because most AI projects fail not on the model but on the data that was not ready, the workflow nobody mapped, or the guardrail nobody specified until the demo went wrong. Related: - [AI readiness for Canadian businesses: a checklist](https://momentaj.com/insights/ai-readiness-checklist-for-canadian-businesses/) - [AI strategy and readiness](https://momentaj.com/services/ai-strategy/) - [How an AI engagement works](https://momentaj.com/how-we-work/#discover) ## Prompt injection URL: https://momentaj.com/glossary/#prompt-injection Prompt injection is an attack in which instructions hidden in something a model reads, such as a document, a web page or a tool result, try to redirect what the model does. A design has to assume it will happen. We test MCP servers with our own injection cases: tool results that contain instructions, descriptions that try to redirect the model, and inputs that smuggle a second request. Related: - [Testing an MCP server](https://momentaj.com/insights/testing-an-mcp-server-what-exists-and-what-is-missing/) - [AI readiness for Canadian businesses: a checklist](https://momentaj.com/insights/ai-readiness-checklist-for-canadian-businesses/) - [MCP servers and AI integrations](https://momentaj.com/services/mcp-development/) ## Evals URL: https://momentaj.com/glossary/#evals Evals (evaluations) are repeatable tests of an AI system against a gold set of real questions or requests with known correct answers, scored before and after every change. For an MCP server, an eval checks that an agent picks the right tool with the right arguments; for retrieval, that the right material comes back. Without one, nobody knows whether a change to a prompt, a model or a description helped. Related: - [Testing an MCP server](https://momentaj.com/insights/testing-an-mcp-server-what-exists-and-what-is-missing/) - [RAG for code and other structured material](https://momentaj.com/insights/rag-for-code-and-structured-material/) ## AI readiness assessment URL: https://momentaj.com/glossary/#ai-readiness-assessment An AI readiness assessment checks whether the workflow, data, rules and people around a process are ready for AI before anyone builds. Ours takes one to two weeks and ends with an opportunity map, the data and guardrail requirements, two or three architecture options with their running costs, a prioritised build plan, and a prototype scope you could hand to any team. It is the first stage of an engagement and can be run on its own. Related: - [AI strategy and readiness](https://momentaj.com/services/ai-strategy/) - [AI readiness for Canadian businesses: a checklist](https://momentaj.com/insights/ai-readiness-checklist-for-canadian-businesses/) - [Notes on AI strategy](https://momentaj.com/insights/tag/ai-strategy/)