Glossary
Glossary
Definitions for agentic AI and sales-analysis terms
A
- Action Spaceفضای اقدام · The action space is the closed list of things an agent is allowed to do. The more open that list, the harder governance becomes.
- Active Customerمشتری فعال · An active customer is a customer who has made at least one purchase in a defined time window. The definition of that window differs by company and must be stated explicitly.
- Agent Evalارزیابی عامل · Agent eval is repeatable measurement of run quality on a known set of questions. Without it, a prompt or model change is only a feeling.
- Agent Handoffتحویل کار بین عاملها · Handoff is the moment one agent passes unfinished work, with the needed context, to the next. Without that context, the next agent guesses from zero.
- Agent Memoryحافظه عامل · Agent memory is what the system keeps from earlier runs so work does not start from zero. Without memory, every question is a session with no history.
- Agent Reflectionبازنگری عامل · Agent reflection is the system scoring its own output against the goal and the evidence before it presents. It is like a critic, but on the same agent’s work.
- Agent Routerمسیریاب عامل · An agent router sends a question to the right template or agent. A campaign-effect question should not enter the customer-churn path.
- Agent Runاجرای عامل · An agent run is one full pass through the pipeline for one question, with an id, time, cost, and result. Observability sits on runs, not on a vague sense of the system.
- Agent Runtimeزمان اجرای عامل · An agent runtime is the environment that hosts the agent loop, tools, and run logging. The model is only one part inside that environment.
- Agent Sessionنشست عامل · An agent session is the time box of one conversation or mission with shared memory. A new session should not attach to the previous one without a reason.
- Agentic AIهوش مصنوعی عاملمحور · Agentic AI is an approach in which the system takes responsibility for completing a task instead of producing an answer. The main difference from conventional AI is the ability to act, not the ability to reply.
- AI Agentعامل هوشمند · An AI agent is a system that, given a specific goal, decides its own steps and uses external tools to get the work done. Unlike a text assistant that answers a question, an agent pursues a task until it produces an output.
- AI Governanceحاکمیت هوش مصنوعی · AI governance is the set of policies and mechanisms that decide what the system is allowed to do, what needs human approval, and how every decision stays reviewable. Without it, organizational use of agentic systems cannot be justified.
- API Connectorاتصال API · An API connector is the bridge between the agent and an outside system: invoices, calendar, messages. A failed connection must be a manageable error for the agent, not a hallucinated success.
- Approval Gateدروازه تأیید · An approval gate is the point where an external action does not proceed without a named human’s sign-off. Proposing is free; executing is not.
- Artifactآرتیفکت · An artifact is a structured, citable output from an agent: not free text, but a record with named fields such as claim, evidence, confidence level, and recommended action. People read the artifact, not the raw run log of the system.
- Audit Logسیاهه ممیزی · An audit log is an immutable record of runs, approvals, and rejects. Traceability without a log is only a claim.
- Autonomy Ladderنردبان خودمختاری · The autonomy ladder is a framework in which each agent’s level of authority rises gradually, based on recorded performance. No agent is allowed independent action from the start.
B
- Baselineخط پایه · A baseline is an estimate of the sales that would have happened without the promotion. Every campaign-effectiveness calculation is measured against this number.
- Business Intelligence (BI)هوش تجاری · Business intelligence is the set of tools that turn organizational data into reports and dashboards. These tools describe the past; they do not explain why it happened and they do not take action.
C
- Cannibalizationکانیبالیزیشن · Cannibalization means part of the discounted product’s sales was taken from another product in your own book. Ignoring it overstates the campaign’s effect.
- Chain-of-Thoughtزنجیره استدلال · Chain-of-thought makes the model write intermediate steps before the final answer. For a manager, those steps must attach to evidence, not to a story.
- Chatbotچتبات · A chatbot is a system that returns a text reply to the user’s message and waits for the next one. Its output is text, and it usually cannot do work in other systems.
- Chunkingقطعهبندی · Chunking splits a document or table into units retrieval can find separately. A chunk that is too large adds noise; one that is too small loses context.
- Circuit Breakerمدارشکن · A circuit breaker stops calling a tool for a while after several failures in a row. That prevents an error cascade and wasted tokens.
- Confidence Levelسطح اطمینان · A confidence level is a statement of how reliable a finding is, together with an explanation of its basis. A confidence level whose basis is not declared has no informational value.
- Context Compactionفشردهسازی زمینه · Context compaction summarizes conversation history or evidence so it fits the context window. The aim is to keep important decisions, not every sentence.
- Context Windowپنجرهٔ زمینه · The context window is the maximum amount of text a language model can consider in one turn. Anything beyond that limit falls out of the model’s view.
- Critic Agentعامل منتقد · A critic agent is the part of the system whose job is to test other agents’ findings and reject those the data does not support. A rejected hypothesis carries as much information as an accepted one.
- Customer Churnریزش مشتری · Customer churn means a customer who was previously active has not purchased in the current window. The churn rate is those customers as a share of all active customers in the prior period.
D
- Data Normalizationیکسانسازی داده · Data normalization means converting different values for the same thing into one standard form. In Persian data this step is necessary, not optional.
- Data Qualityکیفیت داده · Data quality is how complete, accurate, current, and consistent an organization’s data is. No AI system can compensate for weak underlying data.
- Data Residencyمحل نگهداری داده · Data residency is which country or infrastructure stores and processes the data. For an Iranian organization this is an architecture choice, not a cosmetic setting.
E
- Embeddingجاسازی برداری · An embedding is a method for turning text into a set of numbers so that texts with the same meaning have nearby numbers. That makes semantic search possible.
- Episodic Memoryحافظه رویدادی · Episodic memory records one specific run: that session, that question, that result. It is a meeting file, not general brand knowledge.
F
- Fallbackمسیر جایگزین · A fallback is what the system does when the main path fails: a second model, a deterministic query, or a stop with a clear message. A silent failure is the worst case.
- Few-Shotیادگیری با چند نمونه · Few-shot learning puts a few correct examples in the prompt so the model copies the answer shape. It helps artifact format; it does not help computing sales.
- Fine-Tuningتنظیم دقیق · Fine-tuning continues training a model on organization data so tone or format stabilizes. It is not a substitute for retrieval and queries when the truth sits in invoices.
- Function Callingفراخوانی تابع · Function calling is the mechanism by which a model returns a tool’s name and inputs in a structured form so a program can run it. It is the technical layer behind tool use.
G
- Goal Specificationتعریف هدف · Goal specification writes the desired result with a stopping rule, not a vague instruction. “Find out why sales fell” is a goal; “analyze” is not.
- Golden Setمجموعه طلایی · A golden set is questions with a reference answer used to score the agent. It should be built from real manager questions, not from brochure examples.
- Groundingاتصال به شواهد · Grounding means every model claim is tied to a citable record, not to general training knowledge. Without it, a sales figure can sound plausible and still be false.
- Guardrailگاردریل · A guardrail is a written-ahead limit that blocks an output or action outside policy. Unlike a prompt suggestion, a guardrail must be deterministic.
H
- Hallucinationتوهم · A hallucination is a fluent, confident answer that does not match the invoices, SKUs, or campaign records. It comes from how text prediction works: it cannot be removed entirely, but in distribution it can be controlled by tying claims to real data.
- Human-in-the-Loopانسان در حلقه · Human-in-the-loop means the system stops at named points in the process and does not proceed until it receives a human decision. Those points are predefined, not optional.
- Human-on-the-Loopانسان روی حلقه · Human-on-the-loop means the agent proceeds inside an allowed range and a human reviews later or on a sample, not every action. It is a higher rung than human-in-the-loop.
- Human-out-of-the-Loopانسان بیرون حلقه · Human-out-of-the-loop means the agent takes a low-risk action on its own and only reports. In distribution it should be rare and narrow.
- Hybrid Searchجستوجوی ترکیبی · Hybrid search combines semantic search with exact keyword match. A SKU code is not found by meaning; a campaign write-up is not found by keyword alone.
I
- Idempotencyیکباراجرایی · Idempotency means repeating the same request does not create a second effect: invites are not sent twice, budget is not cut twice. Agents need it because retry is normal.
- Incremental Marginحاشیهٔ افزایشی · Incremental margin is the extra net profit from a promotion after subtracting the discount cost on every unit sold. A campaign can raise volume and cut profit at the same time.
- Inferenceاستنباط مدل · Inference is the moment a model runs to produce a reply or a tool request. Cost, latency, and the token ceiling all happen here.
J
K
- Kill Switchکلید قطع · A kill switch stops the whole agent path at once, independent of the model. When errors repeat or cost crosses a ceiling, someone must be able to shut the system off.
- Knowledge Graphگراف دانش · A knowledge graph represents entities and their relations: customer to region, brand to campaign, invoice to SKU. For relational questions it is more precise than text-only search.
L
- Large Language Model (LLM)مدل زبانی بزرگ · A large language model is a system trained on a vast volume of text so that it has learned language patterns and can generate, summarize, translate, and reason in language. It is the engine of many of today’s AI systems, but it is not a product on its own.
- Least Privilegeحداقل دسترسی · Least privilege means each agent and each tool has only the permission its defined job needs. An analysis agent must not be able to change an order.
- Long-Term Memoryحافظه بلندمدت · Long-term memory is what survives between sessions: decision summaries, a manager’s preference, lessons from earlier runs. Writing to it should be deliberate.
M
- Model Cardکارت مدل · A model card is a short document of a model’s limits, training data, and allowed use. Picking a model without reading its limits is not a choice.
- Model Context Protocol (MCP)پروتکل زمینه مدل · MCP is a shared contract for attaching tools and sources to a model so each connection is not written from scratch. A tool is described once and several agents can use it.
- Model Routingمسیریابی مدل · Model routing picks the right model for each job: a small model to classify the question, a strong one to write the agenda. One model for everything is usually waste.
- Multi-Agent Systemسیستم چندعاملی · A multi-agent system is an architecture in which several specialized agents, each with a defined role and scope, work on one problem. That split both raises accuracy and makes error review possible.
- Multi-Step Reasoningاستدلال چندمرحلهای · Multi-step reasoning reaches a result through several inferences in a row, not one instant answer. Each step should be showable.
- Multimodal Modelمدل چندوجهی · A multimodal model understands or produces image or audio as well as text. In distribution that may be a shelf photo or a meeting recording; the claim must still attach to sales data.
O
- Observabilityرصدپذیری · Observability means being able to tell from the outside what the system is doing, how long it takes, how much it costs, and where it fails. For agentic systems, it is a prerequisite for improvement.
- Orchestrationارکستراسیون · Orchestration is the layer that decides which agent runs, in what order, with what input, and where the result goes. Without it, a collection of agents is not a system.
P
- Parallel Tool Callsفراخوانی موازی ابزار · Parallel tool calls run several independent tools at once: brand sales, regional churn, campaign budget. Dependent work must stay serial.
- PII Redactionحذف داده شخصی · PII redaction masks or removes names, phones, and personal identifiers before text is sent to the model. The model does not see what it must not see.
- Pipelineخط پردازش · A pipeline is the sequence of stages a question or a piece of data passes through to reach a final output. Each stage has a defined input, a defined output, and a measurable runtime.
- Planner Agentعامل برنامهریز · A planner agent breaks a goal into runnable steps before any tool is called. The plan names the order of work, not the execution itself.
- Policy Checkبررسی سیاست · A policy check is the evaluation of a proposed action against a set of organizational rules before execution. The result is either pass or stop, and both outcomes are recorded.
- Promptپرامپت · A prompt is the text given to a language model to steer its behavior and output. Prompt quality has a direct effect on answer quality.
- Prompt Injectionتزریق پرامپت · Prompt injection is when input text tries to override the system instruction, for example an invoice or customer email that says “ignore previous rules.” It is an attack on governance, not a normal question.
- Provenanceمنشأ شواهد · Provenance is the path of a figure from the artifact back to the source row. Without provenance, confidence is only tone.
R
- Rate Limitمحدودیت نرخ · A rate limit is a ceiling on requests per unit of time, for the model, a tool, or a user. Without it, a broken loop can explode cost in a minute.
- ReAct Loopحلقه استدلال-اقدام · A ReAct loop is the repeating pattern of thinking, using a tool, seeing the result, and thinking again. The agent does not answer in one shot; it cycles until it has an output.
- Red Teamingتیم قرمز · Red teaming is deliberately testing the system with hostile input and trick questions so a guardrail breaks before production. It is part of eval, not a game.
- Rerankingبازرتبهبندی · Reranking re-orders retrieval results so the more relevant chunk rises. First retrieval is a wide net; reranking tightens it.
- Retrievalبازیابی · Retrieval is selecting a few relevant slices from the organization’s sources before asking the model. RAG without retrieval is only a name.
- Retrieval-Augmented Generation (RAG)بازیابی افزودهٔ تولید · RAG retrieves relevant material from the organization’s own sources and gives it to the language model before the model answers. In distribution, that is how a reply about a SKU, invoice, or campaign is tied to real records instead of the model’s general training.
- Return Rateنرخ مرجوعی · Return rate is returned goods as a share of all goods sold, and it can be measured by quantity or by value. These two numbers are usually different, and each says something different.
- Revenue Concentrationتمرکز درآمد · Revenue concentration is the share of a few large customers in the company’s total revenue. High concentration means losing one customer has a disproportionate effect on the business.
- Run Latencyتأخیر اجرا · Run latency is the gap from question to a usable answer. For a morning sales-manager question, a delay of minutes is no better than a monthly dashboard.
S
- Sales Upliftافزایش واقعی فروش · Sales uplift is the difference between campaign-period sales and the baseline, relative to the baseline. It differs from simple growth in that it compares against the hypothetical “if there had been no campaign,” not against the prior period.
- Sampling Temperatureدمای نمونهبرداری · Sampling temperature is how much random variety the model puts in its output. For numbers and claims it should be low; for a prose draft it can be higher.
- Sandboxمحیط ایزوله · A sandbox is where a tool runs without open access to the whole network or disk. An error or an injection stays in that box.
- Scratchpadپیشنویس میانی · A scratchpad is where an agent writes temporary calculations and hypotheses before the final output. That writing is not for the manager.
- Semantic Searchجستوجوی معنایی · Semantic search means finding results by the meaning of the question, not merely by matching words. The user can ask in natural language without knowing the exact term stored in the data.
- Short-Term Memoryحافظه کوتاهمدت · Short-term memory is the context of the current session: this question, this run’s evidence, this meeting’s decision. It clears or compact when the session ends.
- Small Language Model (SLM)مدل زبانی کوچک · A small language model is cheaper and faster, and enough for a bounded job such as classifying a question or extracting a field. Not every agent job needs a large model.
- State Machineماشین حالت · A state machine limits an agent run to named states: receive, analyze, critique, policy, present. An illegal jump cannot happen.
- Streamingجریان خروجی · Streaming shows tokens gradually instead of all at once at the end. It feels faster, but it must not be the basis of a decision before it finishes.
- Structured Outputخروجی ساختاریافته · Structured output is a reply with named fields, not a free paragraph. Claim, evidence, and recommended action become reviewable once they occupy separate fields.
- Subagentزیرعامل · A subagent is a short-lived agent a supervisor starts for one named job and then shuts down. Its scope and permissions must be tighter than the parent’s.
- Supervisor Agentعامل ناظر · A supervisor agent splits work across specialist agents and collects the result. It does not have to analyze; it coordinates.
- System Promptپرامپت سیستم · A system prompt is the hidden standing instruction that sets a model’s role, limits, and output shape before the user speaks. The user does not see it, but it shapes every reply.
T
- Task Decompositionتجزیه کار · Task decomposition breaks a large goal into smaller jobs with named inputs and outputs. Without it, a manager’s question collapses into a one-shot guess.
- Timeoutمهلت اجرا · A timeout is a time ceiling on the model or a tool. Without it, a stuck query blocks the whole pipeline.
- Tokenتوکن · A token is the smallest unit a language model uses to process text, and it is usually part of a word. Language-model usage is billed by token count.
- Token Budgetبودجه توکن · A token budget is a ceiling on input and output tokens for a run or a day. Without a ceiling, compaction and retrieval lose their point.
- Tool Retryتلاش مجدد ابزار · Tool retry is a controlled repeat of a failed call, with a cap and a gap. Uncapped retry creates cost and side effects.
- Tool Schemaطرح ابزار · A tool schema is the contract for a tool’s inputs and outputs: field names, types, and what is required. The model may only request inside that contract.
- Tool Useاستفاده از ابزار · Tool use is what lets an agent leave description and start doing work: query invoices, write a file, send a message, or book a meeting. In distribution, that is the difference between talking about a campaign and actually reading the ERP.
- Traceabilityردیابیپذیری · Traceability means every system output can be followed back to the data and the stages that produced it. Without it, no decision can be defended in a meeting.
- Trade Marketingبازاریابی تجاری · Trade marketing is the set of activities that target the distribution channel rather than the end consumer: retailer, wholesaler, and distributor. Trade promotions, shelf placement, and channel incentive programs are among its tools.