GLOSSARY
AI Glossary
Technical terms stay English, each with an approved Arabic first-mention gloss.
ANNA search approach that trades a tiny bit of accuracy for huge speed when finding the closest embeddings among millions.AgentA system that uses an LLM to plan, take steps, and call tools toward a goal.Agentic RAGRAG where an agent decides how and when to search — rewriting queries or retrieving multiple times as needed.Approval GateA step that pauses sensitive actions until a human approves them before execution.BackoffIncreasing the wait between each retry (usually exponentially) instead of hammering the server with back-to-back requests.Capability NegotiationThe start of an MCP connection where client and server agree on which features each side supports.ChunkingSplitting documents into smaller pieces before embedding them; piece size and split strategy directly affect retrieval quality.Confused DeputyA flaw where a malicious party abuses a trusted intermediary's permissions (like an agent) to do what it couldn't do itself.Context WindowThe maximum number of tokens a model can see at once.Data ExfiltrationSmuggling sensitive data out of a system — in the agent world, often by tricking the model into sending it.DatasetThe set of examples a model learns from or is evaluated on.EmbeddingA numeric representation of text that places similar meanings close together — the basis of semantic search.EvalA test you run on an AI system with known inputs to measure output quality as a number instead of a gut feeling.Evaluator-OptimizerA pattern where one model generates work and another evaluates it, looping feedback until quality passes.FaithfulnessA RAG metric checking whether the answer is actually supported by the retrieved documents or the model made it up.Fine-tuningAdapting a ready model to your data — an advanced topic rarely needed early.FrameworkA ready-made library that gives you the plumbing for building agents instead of writing everything from scratch.GPUA graphics processor that accelerates training and running models.Golden SetA hand-reviewed set of examples with known-correct answers, used as a fixed reference to measure any change against.GuardrailsConstraints that stop a model from harmful or out-of-scope output.HNSWAn indexing algorithm that makes searching millions of embeddings very fast with approximate accuracy.HallucinationWhen a model states confident but false info because it predicts plausible language, not truth.Hybrid SearchCombining semantic (embedding) search with keyword search to get the best of both.IdempotencyDesigning an operation so running it twice by accident (after a timeout, say) leaves the same result with no duplication.InferenceUsing a trained model to produce output; every API call is inference.JSON SchemaA formal description of the required JSON shape — fields and their types — given to the model so it produces structured output you can rely on.JitterAdding a small random amount to the retry wait so clients don't all retry at the exact same moment.LLMA model trained on huge text to predict the next token — the core of most AI products.Least PrivilegeGive an agent or tool only the minimum permissions it needs, so damage is limited if it's compromised.MCPAn open protocol that standardizes how tools and data connect to AI apps, replacing one-off integrations.MCP ClientThe component inside the host that holds a one-to-one connection with a specific MCP server.MCP HostThe AI application itself (like Claude Desktop) that manages server connections and controls permissions.MCP ServerA program that exposes tools, data, and resources to AI apps via the MCP protocol.ModelThe learned function that turns input into output.Multi-Query RetrievalGenerating several rewordings of the same question and searching with all of them to cover more angles.Multi-hopA question needing several chained retrieval steps, where each step's answer determines the next.OrchestrationManaging the sequence of steps, models, and tools in an AI system so each part runs at the right time.ParametersThe internal numbers a model learns during training and freezes afterward.Parent Document RetrievalSearching with small precise chunks but returning the larger parent document they belong to, so the model gets full context.PrecisionThe share of returned results that are actually correct and relevant.ProbabilisticOutput determined by probability, not certainty — so it varies slightly each run.PromptThe text you give a model to steer its output.Prompt CachingA provider feature that reuses the repeated part of your prompt instead of reprocessing it every call, cutting cost and latency.Prompt InjectionAn attack that hides malicious instructions in content a model reads, making it do things its user never asked.Quality GateA check output must pass before it moves on in the pipeline or reaches the user.Query DecompositionBreaking a complex question into smaller sub-questions answered one by one, then combining the results.Query RoutingSending a user's question to the most suitable source or tool instead of one path for everything.RAGA technique that feeds the model sources at query time so it answers from trusted material and hallucinates less.Rate LimitA cap the API provider sets on requests or tokens per minute; exceed it and you get a 429 error.ReActA pattern where an agent alternates reasoning steps with tool actions, observing each result until it reaches the goal.Reasoning ModelA model that spends tokens on internal thinking before its final answer, solving harder problems at the cost of more time and money.RecallThe share of all relevant items the system actually found out of everything it should have found.ReflectionHaving a model critique its own output and revise it to produce a better version.ReflexionA technique where an agent stores lessons from failed attempts as text memory and uses them on the next try.Regression SetCases that failed before and were fixed, re-run on every change to make sure the problem didn't come back.RerankingPassing initial search results through a more accurate model that reorders them by true relevance.RootsBoundaries the client gives a server to tell it which files or locations it may operate in.RubricAn explicit list of criteria used to judge output quality instead of gut feeling.SSEAn HTTP technique where the server pushes a stream of events over one open connection — the common way LLM responses are streamed.SamplingAn MCP feature letting a server ask the host to run an LLM call on its behalf, with user approval.Self-Query RetrievalThe model converts a user's question into a structured query with metadata filters before searching.StreamingReceiving the model's reply token by token as it's generated instead of waiting for the whole thing.System PromptFixed instructions you set as the developer before any user message, defining the model's role, rules, and style.TemperatureA setting controlling output randomness: lower = steadier, higher = more varied.TokenA small chunk of text (a word or part of one); the unit for length and cost.Tool AggregationThe host gathering tools from multiple MCP servers and presenting them to the model as one list.Tool CallingA model's ability to request an external function/tool in a structured form.Top-pA setting that limits the model to the smallest set of tokens whose probabilities sum to p, trimming off unlikely weird choices.TraceA full record of an agent's steps — every thought, tool call, and result — so you can debug and review.TrainingTeaching a model from data by adjusting parameters — done once and expensive.TransformerThe architecture modern LLMs are built on.