The jargon,
in plain English.

Instead of making things up, the AI looks up the answer in your documents before replying.

Why it matters: this is what separates an assistant that knows your content from a generic one.

The engine. The large language model that writes, summarizes, translates.

Why it matters: not all LLMs are equal — European, American, open, closed. A sovereignty choice to clarify in the contract.

When the AI invents an answer instead of admitting it doesn't know.

Why it matters: this is risk #1. Good RAG + guardrails + source citations reduce it to near zero.

Translation of a text into a vector of numbers so it can be compared to other texts.

This is what lets the AI find "the quote that looks like this one" without being told how.

The unit of measurement. One English word ≈ 1.3 tokens. Models are billed per token.

It's also why short conversations cost less.

Specializing an existing model for a use case. Expensive, slow, rarely necessary.

Beware: a good RAG often beats fine-tuning, at one tenth of the cost.

What you tell the AI. The instruction, the question, the context.

A good assistant has a hidden system prompt that frames tone, scope, guardrails. You don't have to write it.

The limits placed on the AI — forbidden topics, mandatory tone, approved sources.

Without guardrails, the Accounting assistant could start talking politics. Not ideal in front of clients.

The set of documents an assistant is allowed to read.

It's the key concept of isolation at Forgeron3: one client folder = one scope.

The AI answers and indicates where the info comes from (document, page, excerpt).

Without it, you can neither verify nor trust. Always require it.

ANSSI qualification for sovereign clouds. Highest level for sensitive French data.

Essential for local authorities, health data, OIV/OSE operators.

GDPR Article 28 contract between you (controller) and us (processor).

Without a DPA, your compliance is broken. Forgeron3 signs one systematically before go-live.

Impact analysis on data protection. Mandatory for high-risk processing.

Template provided by Forgeron3 for local authorities and sensitive sectors.

Your ability to take back your data and leave, with no friction.

Our commitment: export within 7 days, deletion within 30, open formats, zero exit fees.

Open = accessible weights (Mistral, Llama). Closed = black box billed by usage (GPT-4, Claude).

Forgeron3 uses both depending on the profile — sovereign by default for sensitive data.