Glossary

AI terms · in plain English.

29 terms that come up in real buying conversations — RAG, agents, evals, MCP, guardrails, fine-tuning. Everything is on this one page, so you can search it with Ctrl-F instead of clicking through a tree of stubs.

These definitions are written from the buyer's side rather than the textbook's. Where a term is routinely used to mean two different things — “agent” is the worst offender — the entry says so, because that ambiguity is usually where a scoping conversation goes wrong. Where a concept genuinely needs more than a paragraph, the entry links to the playbook that covers it properly.

A

Agent

An AI system that runs a workflow — fetches data, calls tools, decides between branches, escalates to humans when its confidence is low. Different from a chatbot (which answers questions) in that an agent takes actions.

The defining feature of an agent vs a chatbot is action. Chatbots answer; agents do. In production, agents pair a language model with: (a) a tool-use loop that lets the model call external functions, (b) a memory layer for multi-step context, (c) explicit guardrails on which actions are allowed without human approval, and (d) escalation paths when confidence drops. We design agents exception-first — the refusal cases, escalation queues, and failure modes get scoped before the happy path. See /playbooks/agent-design for the full pattern.

Agentic workflow

A multi-step process where one or more AI agents coordinate to achieve an outcome. Often involves planning, tool use, retrieval, and human-in-loop checkpoints.

B

BAA (Business Associate Agreement)

A contract required under HIPAA when a vendor processes Protected Health Information (PHI) on behalf of a covered entity. We sign BAAs for healthcare engagements and route through enterprise model tiers that include BAAs.

C

Cache (prompt caching)

Reusing the encoded representation of a long, static prompt prefix across many requests. Cuts cost 50-90% on use cases with large shared context (schemas, knowledge bases, instructions).

Anthropic's prompt cache (and OpenAI's structured equivalents) lets you pay for the long static parts of your prompt — system instructions, schemas, large reference docs — once per cache TTL, then ~10% of normal input rate on subsequent reads. For RAG over a large knowledge base with a stable system prompt, this is a 5–10x cost reduction. We surface cache hit rate in observability so it's measurable, not assumed.

Citations / grounding

Requiring the AI to cite the source of each claim it makes from a retrieved document. The single most effective tool against hallucination in RAG systems.

The pattern: numbered chunks in the retrieved context, system prompt requires inline citations like [1] [2] for every factual claim, refusal when no source supports an answer. Hallucination rate drops 80%+ with this pattern in our deployments. See /templates/rag-citation-prompt for the exact prompt we use.

Context window

How much text a model can consider at once. Claude has 200K+ token context windows; useful for working with large documents, codebases, or chat histories.

D

DPA (Data Processing Agreement)

Contract governing how a vendor processes personal data on your behalf. Required under GDPR and India's DPDP Act. We have a template ready or we redline yours.

E

Embedding

A dense numerical representation of text (or other media) that captures meaning. Used for semantic search, clustering, recommendation. Underlies most RAG systems.

Eval / evaluation suite

A set of test cases run weekly against an AI system to catch quality regressions before users do. Non-optional for production. We build one for every shipped system.

20+ test cases at launch, growing weekly with production failures. Each case has expected behavior (not just expected output), scoring rubric, and pass threshold. Run in CI on every prompt change. Failing the eval blocks merge. Named owner on the client side. See /playbooks/eval-suite for the full pattern.

Related: Drift · LLM-as-judge

F

Fine-tuning

Continuing to train a model on your own data to bake in style, vocabulary, or task-specific behavior. Less needed than people think — prompting + RAG usually wins, except for brand voice and visual classification.

Function calling / tool use

When a model decides to call a defined function (search, lookup, write to DB) and the surrounding framework actually executes it. The core mechanism behind agents.

G

Guardrails

Constraints that limit what an agent can do — what files it can touch, what actions need human approval, what topics it refuses. Differentiates production AI from a demo.

H

Hallucination

When a model generates plausible-sounding but false content. Mitigated by RAG with citations, refusal patterns, retrieval-confidence thresholds, and eval coverage. Not zero, never zero.

Hybrid search

Combining keyword search (BM25) and semantic vector search and then reranking the results. Beats either alone on real-world queries. The default for production RAG.

I

Inference

Running a model to get a response. As opposed to training the model. Most LLM cost in production is inference.

L

LLM (Large Language Model)

The big text-prediction models that power everything — Claude, GPT, Llama, etc. Plural is plural, not capitalized LLMs.

LLM-as-judge

Using a strong model to evaluate the outputs of another model against your criteria. Used for eval suites at scale when human grading isn't tractable.

Related: Eval / evaluation suite

M

MCP (Model Context Protocol)

Anthropic's open standard for connecting AI models to tools and data sources. Lets a model call a tool defined in any MCP server. We use it for ERP and internal-systems agents.

Related: Function calling / tool use

Multi-agent system

Multiple specialized agents coordinating to achieve a goal. Mirror-the-org-chart pattern works well — one agent per role you'd hire a human for. See the Multi-Agent Manufacturing case study.

Related: Agent

P

Prompt

The instructions + context you send to a model. Includes a system prompt (fixed, role-defining) and user/assistant messages (the conversation).

Prompt engineering

Crafting prompts to get reliable, high-quality outputs. Real skill. Includes few-shot examples, chain-of-thought, structured output, refusal patterns.

R

RAG (Retrieval-Augmented Generation)

Fetching relevant documents from a database first, then asking the model to answer using only those documents. The default architecture for grounded chatbots and knowledge assistants.

Production RAG isn't just 'fetch chunks and stuff them in the prompt.' It's hybrid search (BM25 + vector) → reranking → citation-required prompting → refusal patterns when retrieval confidence is low. Hallucination drops 80%+ with the full pattern vs naive RAG. See /playbooks/rag for the full pattern.

Related: Embedding · Vector database · Reranking

Reranking

Running retrieved candidates through a second model to reorder by relevance. Top-50 from vector search reranked to top-5 dramatically improves precision. We use Cohere Rerank.

S

SCCs (Standard Contractual Clauses)

EU-approved contract language for transferring personal data outside the EU. Included in our DPA. Required for most EU↔India data flows.

Sandbox

An isolated execution environment for untrusted code or agent actions. Used in our Auto Issue Resolution system to let Claude Code write fixes safely.

Streaming

Returning the model's response token-by-token as it's generated, instead of waiting for the full reply. Critical for chat UX — feels responsive even when the full response takes seconds.

System prompt

Fixed instructions at the top of every conversation — role, tone, allowed behaviors, refusal patterns. The lever you tune most often when shipping production AI.

T

Token

The unit of text models process. Roughly 4 characters or ¾ of a word in English. Costs and rate limits are denominated in tokens.

V

Vector database

A database optimized for similarity search over embeddings. pgvector if you're on Postgres, Pinecone / Qdrant / Weaviate when you need more scale.

The three terms worth arguing about

Most of the vocabulary above is stable. These three aren't, and almost every misaligned AI project we've been asked to rescue traces back to one of them meaning different things to the buyer and the builder.

“Agent”: Vendors use it for anything from a scripted chatbot to a system that takes irreversible actions unsupervised. The distinction that matters commercially is whether it acts: a system that only answers needs a good retrieval layer, while a system that acts needs an action allowlist, approval gates, and an escalation path. Those are different budgets. Settle which one you mean before anyone quotes.
“Fine-tuning”: Frequently proposed when retrieval is the actual answer. Fine-tuning changes how a model behaves; it is a poor way to give a model facts, because the facts go stale and retraining is expensive. If the requirement is “it should know our policies,” that is a RAG problem, not a fine-tuning one.
“Accuracy”: A number with no test set behind it. Ask what it was measured on, how many cases, and who wrote them — if the answer is a demo and a good feeling, there is no accuracy figure, there is a vibe. This is the entire reason an eval suite exists.

The longer versions live in the playbooks — agent design, RAG, and the eval suite respectively.

Term missing? Email sales@aliansoftware.net and we'll add it.

Talk to a human about this →