Glossaire · 69 termes

Le vocabulaire, décodé

Le jargon tech sert souvent à impressionner. Ici, il sert à comprendre : les mots qu'on emploie avec vous, expliqués simplement — pour que vous décidiez en connaissance de cause.

01 — Intelligence artificielle

Intelligence artificielle

23 termes.

Computer Vision: AI branch enabling computers to interpret and analyze images and videos. Applications: object detection, facial recognition, quality control.
LLMLarge Language Model: Artificial intelligence model trained on massive amounts of text, capable of understanding and generating natural language. Examples: GPT, Claude, Mistral.
NLPNatural Language Processing: AI field dealing with interaction between computers and human language. Enables text analysis, translation, generation, and contextual understanding.
OCROptical Character Recognition: Technology converting printed or handwritten text images into editable digital text. Essential for document automation.
Prompt InjectionPrompt Injection Attack: Attack technique against AI systems that inserts malicious instructions into user inputs to hijack an AI agent's behavior. Can cause data exfiltration or unauthorized actions.
RAGRetrieval-Augmented Generation: A technique combining retrieval of information from a knowledge base with generation by an LLM. For each query, the system retrieves the relevant passages (ideally via hybrid search: vector + BM25, then reranking) and supplies them to the model as context — accurate, sourced, up-to-date answers without retraining the model.
EmbeddingSemantic vector: A numerical representation of a text (or image) as a vector capturing its meaning. Enables measuring semantic similarity between content — the basis of vector search and RAG.
ChunkingDocument segmentation: Splitting a document into manageable segments (with overlap) before indexing. Good chunking preserves context and improves retrieval precision in a RAG system.
Hybrid searchHybrid Search: Combining keyword search (BM25) with vector (semantic) search, fused via a method such as RRF. Captures both exact terms (codes, references) and meaning, where vector-only search fails.
RerankingResult reranking: A second relevance pass over retrieved results: a cross-encoder model re-scores each passage against the query to keep only the most relevant ones before generation. A major precision boost for a RAG system.
Context engineeringContext Engineering: The discipline of assembling exactly the right information into the model's context window at each step (system prompt, retrieved data, tools, memory, history). The system-level evolution of prompt engineering.
AI agentAgentic system: A system where an LLM plans, calls tools and acts over multiple steps toward a goal, rather than producing a single answer. Deployed autonomous, guided, or human-supervised.
MCPModel Context Protocol: An open standard connecting AI models to tools and data sources through reusable servers, instead of bespoke per-system integrations. Think of it as 'USB-C for AI tools'.
A2AAgent-to-Agent: A protocol letting autonomous AI agents communicate and delegate tasks interoperably. Complementary to MCP: MCP is for calling a tool, A2A for delegating to another agent.
Fine-tuningModel fine-tuning: Targeted retraining of a model on specific data (often via LoRA/QLoRA) to adapt its style or domain. Complements RAG rather than replacing it: RAG provides knowledge, fine-tuning provides behavior.
AI evaluationEval / RAGAS: Systematic measurement of an AI system's quality (faithfulness, relevance, hallucination rate, retrieval quality) via automated test sets, ideally wired into CI. Measure quality, don't assume it.
HallucinationAI hallucination: When a model generates plausible but false information unsupported by the provided sources. RAG, source grounding and faithfulness evaluation reduce it — they never fully eliminate it.
AI memoryAgent memory: A mechanism letting an assistant or agent retain and recall information across sessions (preferences, facts, history), beyond the context window. Organized in tiers: short-term and long-term.
Multi-tier memoryMulti-level AI memory: Organising an AI system's memory across several levels — immediate (the conversation), episodic (past sessions), long-term (consolidated knowledge) — to retain what matters without keeping everything. This is what lets an assistant pick up where you left off.
Knowledge graph: A representation of knowledge as connected entities (clients, contracts, projects — who, what, when). It lets an AI answer questions that span multiple sources instead of merely retrieving an isolated paragraph (multi-hop retrieval, GraphRAG).
Retrieval evaluation: Automated tests that continuously measure whether a RAG system retrieves the right information: answer faithfulness to sources, relevance of retrieved passages, context precision. Without it, the quality of an enterprise AI drifts silently.
Guardrails: Input and output controls on an AI system that filter unwanted content, data leaks and prompt-injection attempts. Essential for reliable production AI.
Context windowContext Window: The maximum amount of text (measured in tokens) a model can process in a single request, input and output combined. Managing it well is decisive for cost, latency and quality.

Ces briques, on les met en production :

Cœur IA d'entreprise Agents IA Diagnostic IA

02 — Développement

Développement de A à Z

16 termes.

APIApplication Programming Interface: Set of rules and protocols enabling different software applications to communicate with each other. Essential for system integration and data exchange between applications.
CRUDCreate, Read, Update, Delete: Four basic operations for data manipulation in an application. Fundamental for any data management interface.
DockerContainerization Platform: Technology enabling application and dependency packaging in standardized containers. Ensures portability and reproducibility across environments.
ETLExtract, Transform, Load: Data integration process from multiple sources to a target system. Essential for enterprise data consolidation and analysis.
GraphQLGraph Query Language: Query language for APIs allowing clients to request exactly the data they need. Flexible alternative to REST, developed by Facebook.
ISRIncremental Static Regeneration: Next.js technique allowing static page updates after build without full rebuild. Combines SSG and SSR benefits.
KubernetesContainer Orchestration System: Open-source platform automating deployment, scaling, and management of containerized applications. Industry standard for cloud native.
Multi-tenantMulti-tenant Architecture: Software architecture where a single application instance serves multiple clients (tenants) with complete data isolation. Standard for SaaS platforms.
ORMObject-Relational Mapping: Programming technique converting data between objects in code and database tables. Simplifies data access.
PWAProgressive Web App: Web application offering native-like experience: offline functionality, push notifications, home screen installation.
RESTRepresentational State Transfer: Standard architecture for designing web APIs. Uses HTTP methods (GET, POST, PUT, DELETE) to manipulate resources uniformly.
SPASingle Page Application: Web application loading a single HTML page and dynamically updating content. Offers smooth experience similar to desktop application.
SSGStatic Site Generation: Method of generating HTML pages in advance during build time. Offers maximum performance and excellent security.
SSRServer-Side Rendering: Technique of rendering web pages on the server before sending to browser. Improves SEO and perceived performance.
TerraformMulti-cloud IaC Tool: Open-source tool for defining and provisioning cloud infrastructure declaratively. Compatible with AWS, GCP, Azure, and many other providers.
WebhookAutomated HTTP Callback: Mechanism allowing an application to automatically send real-time data to another application upon specific events.

03 — Cloud & SaaS

Cloud & SaaS

4 termes.

AWSAmazon Web Services: World-leading cloud platform offering on-demand computing, storage, databases, and infrastructure services. Used by millions of businesses worldwide.
CDNContent Delivery Network: Geographically distributed network of servers for fast web content delivery. Reduces loading times and improves overall performance.
GCPGoogle Cloud Platform: Google's suite of cloud services including computing, storage, Big Data, machine learning, and developer tools.
SaaSSoftware as a Service: Software distribution model where the application is hosted in the cloud and accessed via the internet through subscription. Eliminates the need for local installation.

04 — Méthode & DevOps

Méthode & DevOps

6 termes.

CI/CDContinuous Integration / Continuous Deployment: DevOps practice automating code testing and deployment. Enables rapid and confident delivery of new features.
DevOpsDevelopment and Operations: Practices combining software development and IT operations to shorten development cycles and continuously deliver quality code.
IaCInfrastructure as Code: Practice of managing infrastructure through versioned code. Enables automated and reliable reproduction, testing, and deployment of infrastructures.
ScrumScrum Agile Framework: Agile project management framework structured in short sprints (1-4 weeks). Promotes collaboration, adaptation, and incremental delivery.
SprintDevelopment Iteration: Fixed time period (1-4 weeks) during which an agile team delivers a functional product increment. Regular development rhythm.
TDDTest-Driven Development: Development practice where tests are written before code. Ensures complete test coverage and more robust code.

05 — Sécurité

Sécurité de A à Z

12 termes.

CVECommon Vulnerabilities and Exposures: International reference system for cybersecurity vulnerabilities. Each CVE is a unique identifier (e.g., CVE-2026-25253) assigned to a known flaw, enabling standardized tracking and remediation worldwide.
CVSSCommon Vulnerability Scoring System: Standardized scoring system evaluating the severity of security vulnerabilities on a scale from 0 to 10. A score of 8.8 or higher indicates a critical vulnerability requiring immediate remediation.
ISO 27001Information Security Management Standard: International standard defining requirements for information security management system (ISMS). Ensures sensitive data protection.
OAuthOpen Authorization: Standard protocol for secure authentication and authorization. Enables SSO (Single Sign-On) and access delegation between applications.
OWASPOpen Web Application Security Project: Nonprofit organization publishing the most critical web security vulnerabilities. OWASP Top 10 is the global reference for application security.
RBACRole-Based Access Control: System access restriction method based on user roles. Simplifies permission management in multi-user applications.
RCERemote Code Execution: Type of vulnerability allowing an attacker to execute arbitrary code on a target system without physical access. Among the most dangerous flaws in cybersecurity.
GDPRGeneral Data Protection Regulation: European regulation governing personal data collection and processing. Imposes strict obligations with fines up to 4% of global revenue.
SandboxingSandbox Isolation: Security technique that runs a program in an isolated environment (sandbox) to limit its access to system resources. Prevents a compromised component from affecting the rest of the infrastructure.
SOC 2Service Organization Control 2: Security certification for cloud service providers. Attests to security controls, availability, confidentiality, and compliance.
XSSCross-Site Scripting: Vulnerability allowing malicious code injection in web pages viewed by other users. Can steal sensitive data or hijack sessions.
SQL Injection: Attack technique exploiting security flaws to execute malicious SQL code. Can enable data theft, deletion, or unauthorized modification.

06 — Produit

Produit de A à Z

5 termes.

MVPMinimum Viable Product: Simplified version of a product with just enough features to satisfy early adopters and validate a market hypothesis.
POCProof of Concept: Experimental realization to demonstrate feasibility of an idea or technology. Validates approach before full investment.
UIUser Interface: Visual part of an application with which the user interacts. Includes buttons, menus, forms, and all graphical elements.
UXUser Experience: Set of emotions and perceptions of a user when interacting with a product. Encompasses usability, accessibility, and satisfaction.
ASOApp Store Optimization: Set of techniques aiming to improve mobile application visibility and ranking in stores (App Store, Google Play).

07 — Business

Business de A à Z

3 termes.

CTOChief Technology Officer: Executive responsible for a company's technology strategy. Defines architecture, manages technical team, and aligns tech with business objectives.
KYCKnow Your Customer: Customer identity verification process, mandatory in financial sectors. Prevents fraud, money laundering, and terrorism financing.
ROIReturn on Investment: Measure of investment profitability, expressed as ratio between net gain and initial cost. Essential for evaluating project relevance.

MyoApp · La Rochelle

Un terme qui reste flou ?

On vous l'explique de vive voix, appliqué à votre contexte — une conversation, pas un cours magistral.

Parler de votre projet

On écoute vraiment. Le reste suit.