What is an LLM? RAG and OCR: AI terms explained for SMEs

LLM stands for Large Language Model (not to be confused with the LL.M., a master’s degree in law). It is a program trained on very large amounts of text, which writes by predicting, piece by piece, the most likely continuation.

ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google) and Le Chat (Mistral AI) run on LLMs. An example in an SME: you paste in a complaint e-mail, the LLM suggests a polite reply, and you read it over before sending.

Lumineth uses Claude, from Anthropic, for analysis and writing. The LLM proposes, the code checks every figure, and a person accepts or rejects.

The prompt is the text the LLM receives: the instruction, the context and often an example of the expected result. The more precise the instruction, the more usable the answer.

Example: “Summarise this customer e-mail in two sentences, then suggest a reply in a formal tone, without promising a deadline.” This instruction gives more consistent results than “Reply to this customer”.

In a tool running in production, a prompt is written and tested like code. At Lumineth, every prompt is versioned, its fingerprint is logged on each run, and any change to a prompt is first replayed on a set of reference cases.

RAG stands for Retrieval-Augmented Generation. Before answering, the system searches your documents for the relevant passages, then the LLM writes its answer from those passages and cites the source.

Example: a heating company keeps its technical sheets in a shared folder. A technician asks what an error code means; the assistant finds the right sheet and answers with a link to it.

In Lumineth Hub, the “Help” button works this way: a search through the guides, then Claude Haiku 4.5 answers from the passages found, with sources cited. The principle is explained in detail in the guide to AI knowledge bases (RAG).

OCR stands for Optical Character Recognition. A scanned PDF or a photo of a document is only an image: OCR pulls the text out of it, so it can then be searched, copied or read by an LLM.

Example: supplier invoices arrive as scanned PDFs. OCR reads the text, then the supplier, date and amount are extracted and prepared for the accounts; a person checks them before approval.

Lumineth processes mail, letters and scanned invoices this way, as well as documents photographed on a phone. When a reading is uncertain, it goes to a person, who corrects it: the process is described on the processes and documents page.

An AI agent takes on a repetitive task, with an LLM for the part that calls for judgement: reading an e-mail, sorting a request, preparing a reply. A chatbot answers in a chat window; an agent works inside your tools.

Example: an SME receives requests by e-mail every day. The agent reads them, links each one to the right customer file and prepares a draft reply, which you read over and send.

At Lumineth, the path an agent follows is set by the code: the code calculates, the model proposes, the code checks, a person decides. For an assistant that answers your customers, see the article on AI chatbot pricing and how it works.

An API is an entry point that a piece of software opens to other programs; many CRMs, accounting packages and calendars offer one. AI connects to it to read or write data, and you keep the tools you already use.

Example: once the invoice has been read, the entry is created in your business software through its API, with no manual re-keying.

MCP (Model Context Protocol) is an open standard, published by Anthropic in 2024, that lets an AI assistant use external tools. In Lumineth Hub, an MCP server lets your AI assistants read your data with your access rights, and your business tools connect through an API, with signed exchanges.

An LLM predicts the most likely text without checking anything on its own. When it lacks the information, it can write a well-phrased answer that is wrong, such as an invented deadline or figure: this is called a hallucination.

Example: asked about your returns policy without access to your documents, an assistant may state a return period you have never offered.

The safeguards work together: RAG so that answers come from your sources, checks written in code, and a person when the answer is uncertain. At Lumineth, any number the model writes must appear in the facts collected, otherwise the proposal is blocked.

AI prepares, you decide. You choose which task to hand over, what stays in your hands and the confidence threshold below which a case comes back to you.

Example: a draft reply, an invoice that has been read or a sorted request waits for your approval, and nothing goes out without your consent. Each correction serves as an example and improves what follows.

Decisions are recorded in a log, and no personal data is sent to the model without your consent. The full framework is on the AI agents in Geneva page.