Artificial intelligence brings its own vocabulary, which can look opaque. This glossary is here to help you read BrainDup's documentation, along with our Development and Contact pages, without a degree in computer science. Every term is explained in plain language, with a concrete comparison where it helps.
Infrastructure and hardware
The physical and software parts that run BrainDup on your own machine, without the cloud.
GPU
A graphics processor, originally built for display, today indispensable for AI. Where a classic processor works through tasks one at a time, the GPU handles thousands at once, like an army of parallel calculators.
NVIDIA Blackwell
NVIDIA's latest GPU architecture (2024-2025). The Blackwell cards (RTX 50xx series) deliver markedly more AI performance than earlier generations, at lower power draw.
VRAM
The GPU's own memory. The more VRAM, the larger the AI model the GPU can load in one go. It is the main limit on running language models locally.
CUDA / cuDNN
NVIDIA's software toolkit for tapping GPU power. CUDA is the general base; cuDNN is the layer specialised for AI and deep learning.
Driver
Software that lets the operating system talk to hardware, here the GPU. Without the right driver the GPU is unusable. BrainDup keeps the NVIDIA drivers current.
Docker
A tool that packs an application with everything it needs (libraries, configuration) into a sealed bundle called a container. Think of an airtight box: the application behaves identically anywhere.
Container
An isolated, lightweight environment created by Docker. Each BrainDup building block (database, AI model, interface) runs in its own container, which avoids conflicts and simplifies updates.
Orchestration
Automatic coordination of several containers that must work together. BrainDup uses Docker Compose to start, stop and connect its services with a single command.
NGINX
A very fast web server used as the single entry point to BrainDup's services. It spreads the traffic, handles HTTPS security and shields the internal services.
Reverse proxy
An intermediary placed in front of your web services. It takes every outside request and routes it to the right internal service, like a switchboard directing calls.
HTTPS
The secure communication protocol of the web. The padlock in your browser shows the exchanged data is encrypted. BrainDup enforces HTTPS on all traffic.
Server
A computer (or program) providing services to other machines. With BrainDup, your local machine acts as your own server: it holds your AI models and your data.
Local (on premises)
Said of infrastructure installed physically in your own building, as opposed to the cloud. With BrainDup, “local” means your data never leaves the building.
Cloud
A set of remote servers (data centres) rented from third parties (Amazon, Google, Microsoft). Convenient, expensive over time, and your data sits outside your control.
Data sovereignty
An organisation's ability to keep complete control of its data: where it is stored, who reaches it, under which jurisdiction. It is BrainDup's founding principle.
Databases and storage
How BrainDup stores, indexes and retrieves your documents and knowledge.
Database
An organised system for storing and retrieving information. Like a clever digital filing cabinet handling millions of cards and answering complex queries in a fraction of a second.
PostgreSQL
The reference open source relational database, used by BrainDup for structured data (users, projects, history). Reliable, fast and free.
Milvus
A database specialised in storing vectors. It lets BrainDup retrieve the most relevant documents by comparing their “semantic fingerprint” rather than keywords.
Vector database
A type of database storing mathematical representations (vectors) of the meaning of texts. It finds similar documents even when they use different words.
Vector (embedding)
A list of numbers representing the “meaning” of a text in a mathematical space. Two texts close in meaning get close vectors, even with different words.
Embedding
The process of turning a text (or an image) into a numerical vector. It is the first step of RAG: every document is converted to enter the vector database.
Semantic search
Search based on meaning rather than exact words. Searching for “staff remuneration” also finds a document about “wages and salaries”.
Neo4j
The graph database BrainDup uses to map relationships between concepts: who works with whom, which document refers to which other.
Knowledge graph
Knowledge represented as a network: nodes (concepts) joined by typed relationships. It allows questions no single document answers literally.
MinIO
An open source file storage server, compatible with the S3 standard. BrainDup uses it to keep documents and derived files locally.
S3
A file storage standard invented by Amazon (Simple Storage Service). MinIO speaks the same language, so the tooling works without any cloud dependency.
Nextcloud
An open source alternative to Google Drive or Dropbox. It syncs, shares and co-edits files on your own server.
Collabora
An office suite in the browser (text, spreadsheet, presentation), integrated with Nextcloud. A sovereign alternative to Office 365 or Google Docs.
Artificial intelligence and language models
The concepts behind the models BrainDup runs locally.
Artificial intelligence (AI)
A set of techniques letting a machine imitate human abilities: understanding language, recognising images, deciding, drafting text.
Model (an AI one)
A program trained on large amounts of data to perform a given task (drafting text, recognising an image). The model is the result of that training.
LLM
Large Language Model, a language model of large size, able to understand and produce language. Mistral, Llama and Claude are LLMs.
Inference
The phase where the AI model produces an answer from a question. It is the moment the GPU computes; training happened earlier.
Model parameters
Numerical values, adjusted during training, that determine the model's behaviour. A 7-billion-parameter model has seven billion of them.
Mistral
A French AI company building capable open source language models. BrainDup uses them for local work in French.
Qwen
A family of open source AI models from Alibaba. Excellent performance for their size, and genuinely multilingual.
Claude
An AI model from Anthropic, known for its accuracy and for following instructions. AS3P uses it to develop, through Claude Code.
ChatGPT / GPT
OpenAI's AI models, the best known to the general public. Powerful, but cloud-only: your data leaves your building.
Gemini
Google's AI model. Like GPT it performs well, but it runs only in the cloud, with the same sovereignty question.
Ollama
An open source tool that simplifies downloading and running AI models on your own machine. It is the engine of BrainDup's local AI.
Vision model / VLM
A model that understands both text and images. Useful for reading scanned documents, plans or photographs.
OCR
Optical character recognition: technology that turns an image of text (a scan, a photo) into text a machine can process.
Hallucination
When an AI model invents false information and presents it with confidence. RAG cuts that risk by tying the model to real documents.
RAG
Retrieval-Augmented Generation: first find the relevant documents, then let the model answer from them. It is the core of BrainDup.
Fine-tuning
Partially retraining an AI model on your own data to specialise it. Like upskilling an experienced colleague rather than hiring a new one.
Prompt
The instruction or question you give an AI model. The quality of the prompt strongly shapes the quality of the answer.
Token
The basic unit of text for a language model. One token is roughly three quarters of a word. Cloud services bill per token; locally it costs nothing.
Context
The amount of text a model can take into account at once to form its answer. A larger context window allows long documents.
Open source
Software whose code is public, free and modifiable. BrainDup favours open components to avoid vendor lock-in.
Hugging Face
A community platform where thousands of open source AI models are shared freely. It is where BrainDup gets its models.
Software development and code
The vocabulary you will meet in BrainDup's technical documentation.
Python
The reference programming language for AI and data processing. Simple, readable, with a rich set of libraries. Most of BrainDup is written in Python.
Code
Instructions written in a programming language telling the computer what to do. BrainDup's source code is public and auditable.
FastAPI
A modern Python framework for building fast web interfaces that document themselves. BrainDup uses it so its services can talk to each other.
API
An application programming interface: a standardised contract letting two programs communicate. Like a standardised socket: any compatible device plugs in.
Endpoint
The precise address of an API where a request goes for a specific service. For example, an endpoint to ingest a document.
JSON
A universal format for exchanging data between applications, readable by humans and machines. It is the lingua franca of web interfaces.
Pipeline
A chain of automated steps run in a set order. In BrainDup, a RAG pipeline strings together: receive the document, extract the text, chunk it, vectorise, store.
Unit test
An automated check that a small piece of code (a function) works correctly in isolation. BrainDup contains hundreds of them.
Integration test
A check that several components work correctly together. Having tested each piece separately, you test the chain.
Regression
A bug introduced unintentionally while changing existing code. A regression test verifies that earlier fixes still hold.
Bug
An error in a program causing unexpected behaviour. Finding and fixing bugs, debugging, is a standing part of the work.
Framework
A development frame providing structure and ready-made tooling. Rather than building everything yourself, you start from a shared base.
Library
A collection of reusable functions written by others. Like borrowing a book rather than writing it yourself.
Version control and releases
How BrainDup's code evolves, gets reviewed and ships.
Git
A version control system recording every change to the source code, like a complete history you can step back through.
GitHub
An online platform hosting Git repositories. It eases collaboration between developers and offers tooling for automation.
Commit
Recording a set of changes in Git, with a descriptive message. Like saving, but with the reason attached.
Branch
A parallel copy of the code, for building a feature without touching the stable version. Once ready, it is merged.
Pull request
A formal request to merge a branch into the main code. It lets other developers review before it becomes final.
CI/CD
Continuous integration and continuous delivery: tests and deployment are automated. Every change goes through the same pipeline.
Release / Version
The official publication of a version, identified by a number (for example v2.1.0). Each release bundles the changes since the previous one.
Orchestration, workflows and integration
How BrainDup makes its own parts and your software work together.
Workflow
A sequence of automated steps triggered by an event. For example: a document arrives, it is analysed, then stored.
n8n
An open source automation platform BrainDup uses to build workflows visually, without code.
Node
In n8n, each step of a workflow is a node: receive a file, call the AI, send an email. Nodes are chained together.
Webhook
A mechanism letting one application notify another automatically when something happens. No more constant polling.
Batch
Processing several items together in a single operation. Instead of 100 documents one by one, they are handled as a group.
Retry
An automatic further attempt after a failure. If a step fails (unstable network, service briefly down), it is tried again.
Fallback
A backup route activated when the primary one fails. If the main AI model is overloaded, for instance, another takes over.
Method and project governance
The concepts that frame how BrainDup is built and kept compliant.
BMAD
A method for running AI projects, used to structure BrainDup's development. It sets the roles, the phases and the deliverables.
Story
A short description of a feature from the user's point of view. For example: “As a lawyer, I want to search the deeds in a case file.”
Sprint
A short, fixed working period (usually one to two weeks) in which the team completes an agreed set of stories.
Backlog
An ordered list of every feature, improvement and fix still to be done. The backlog is alive: it is reordered as priorities move.
PRD
Product Requirements Document: the document describing precisely what a product must do, for whom and why.
GDPR
The General Data Protection Regulation, the European law imposing strict rules on collecting and processing personal data.
EU AI Act
The European regulation on artificial intelligence, in force since 2024. It classes AI systems by level of risk and attaches obligations to each.
Technical debt
An accumulation of shortcuts and compromises that slow future development. Like a loan: you gain time now and repay with interest later.
Alphabetical index
Take the step
Thirty minutes to assess what sovereign AI can do in your organisation. No commitment.
Free sovereignty review