AI Glossary

A set of techniques enabling machines to perform tasks that normally require human intelligence: language understanding, pattern recognition, decision-making. Modern AI is based on machine learning.

A large-scale language model trained on vast text corpora. LLMs understand and generate natural language with remarkable quality. Examples: GPT-4o (OpenAI), Claude 3.5 (Anthropic), Llama 3 (Meta), Mistral (Mistral AI).

An architecture where the LLM is augmented by a search in an external document database. Rather than answering from its training memory, the model consults your specific documents and generates a response based on those sources. This enables precise, traceable and up-to-date answers.

An instruction or question submitted to an LLM to obtain a response. The quality of the prompt directly influences the quality of the response. Prompt engineering is the art of formulating optimal instructions to get the best behaviour from a model.

The processing unit of an LLM. One token corresponds approximately to 3/4 of a word in English. LLMs have a limited context window (e.g. 128,000 tokens for GPT-4o) that determines how much text they can process in a single request.

The maximum amount of text an LLM can process in a single interaction, measured in tokens. The larger the window, the more text the model can analyse simultaneously. GPT-4o supports 128,000 tokens (~96,000 words), roughly 200 pages.

A phenomenon where an LLM generates a plausible but factually incorrect response. It is one of the main challenges of generative AI. SkaLean's RAG pipeline mitigates this risk by grounding every response in your verified source documents.

The process of re-training a pre-trained LLM on a specific dataset to improve its performance in a particular domain. Allows the model to adapt to specialized terminology, a writing style or precise business rules.

The process by which an AI model generates a prediction or response from an input. Requires significant computing power. SkaLean runs inference on local sovereign GPUs to ensure your data never leaves your country.

An autonomous AI system capable of planning and executing a sequence of actions to achieve a goal. Unlike a simple chatbot, an agent can call tools, consult external data, and chain multiple reasoning steps (ReAct loop).

A mathematical representation of the semantic meaning of a text as a high-dimensional numerical vector. Two texts with similar meaning will have similar vectors. Embeddings are the foundation of semantic searches in the RAG pipeline.

The splitting of a document into optimally sized fragments (chunks) before indexing in the vector database. Smart chunking preserves context: not too small (loses meaning) nor too large (dilutes precision). SkaLean uses adaptive semantic chunking.

A parameter that controls the creativity of an LLM's responses. Temperature 0 = deterministic and precise responses (ideal for factual tasks and RAG). Temperature 1 = varied and creative responses. SkaLean uses low temperatures for business use cases.

The discipline of designing optimal instructions to get the best behaviour from an LLM. Includes writing the system prompt, structuring examples and managing constraints. SkaLean includes optimized business prompt templates.

A base instruction provided to the LLM before any user interaction. Defines the AI's role, tone, rules of conduct and constraints. In SkaLean, each agent has a customizable system prompt to reflect your brand and policies.

An automatic process that transforms a raw document into a cited, sourced AI answer. SkaLean combines ingestion, semantic preprocessing, vector indexing and LLM synthesis to retrieve the right information from your documents, even when phrased differently from your question.

A specialized database storing embedding vectors and enabling search by semantic similarity (nearest neighbors). Unlike SQL databases, vector search retrieves semantically similar documents even if the exact words differ.

The semantic storage and search infrastructure used by SkaLean to index your documents. Supports millions of vectors, metadata filtering and hybrid search (semantic + keywords), ensuring precise, sourced AI responses.

A high-performance software layer enabling AI models to run with minimal latency and high throughput. SkaLean uses its own sovereign GPUs to run models like Llama 3 or Mistral, serving multiple simultaneous requests without dependency on third-party APIs.

Optimization of LLM inference at the hardware level (GPU) for speed gains of 2 to 4x compared to standard inference. SkaLean deploys this optimization in production to guarantee response times under 200ms.

An AI model fine-tuned on an organization's specific business data (medical, legal, financial). SkaLean offers sector-specific fine-tuning of open-source models to adapt vocabulary, tone and knowledge to your sector's precise needs.

A mechanism allowing multiple users to work simultaneously on the same document or workflow without conflicts or modification overwrites. SkaLean integrates real-time collaboration in AI Studio for co-editing documents and knowledge bases.

A PostgreSQL security mechanism restricting access to table rows based on the user's identity. SkaLean uses RLS for multi-tenant isolation: each client can only access their own data, even if stored in the same database.

An architecture where multiple clients (tenants) share the same infrastructure while having their data isolated. SkaLean is multi-tenant with database-level isolation (RLS): clients share GPU resources and infrastructure costs, but their data is strictly separated.

A service-to-service communication architecture based on real-time event streams. SkaLean uses event streaming for communication between its modules: authentication, messages, automation executions, agent interactions, etc. Guarantees reliable event delivery.

A bidirectional real-time communication protocol between a browser and a server. SkaLean uses WebSockets for real-time AI chat, automation status notifications and real-time document collaboration.

A mechanism allowing an application to send an HTTP notification to another application when an event occurs. SkaLean exposes webhooks for all key events: message received, workflow executed, agent escalation. Enables integration with third-party systems.

The principle that an organization's data must remain under the legal jurisdiction and physical control of its home country. SkaLean guarantees sovereignty by hosting data and AI computations on servers physically located in your country.

A set of laws governing the collection, use and sharing of personal data. Imposes obligations including data residency, right to portability and incident notification. Examples: GDPR (Europe), CCPA (USA), and applicable local laws in your territory.

General Data Protection Regulation. European regulation on the protection of personal data, also applicable to organizations outside the EU processing data of European residents. SkaLean is GDPR-compliant for its European clients.

Health Insurance Portability and Accountability Act. US law on the protection of health data. SkaLean meets all 17 HIPAA criteria for medical sector clients: access control, audit trail, encryption, recovery plan, BAA agreement.

A list of the 10 most critical security risks specific to LLM-based applications, published by the Open Web Application Security Project. SkaLean protects against 10/10 of these risks: prompt injection, data leakage, hallucination, denial of service, etc.

An attack where a malicious user inserts instructions into a message to manipulate an LLM's behaviour (e.g. ignoring security rules, revealing confidential information). SkaLean implements multiple protection layers against these attacks.

Personal data that can directly or indirectly identify an individual: name, address, social security number, credit card number, etc. SkaLean automatically detects and masks 15 types of PII before sending to the LLM.

A bank-grade symmetric encryption standard used by SkaLean for data at rest. The 256-bit key is considered unbreakable with current technology, and the authentication mode guarantees data integrity against any tampering.

A security protocol encrypting network communications. SkaLean uses TLS 1.3 (the latest version) for all communications in transit: between browser and server, between internal services, and between SkaLean infrastructure and external LLM APIs.

An identity verification mechanism combining multiple factors: something you know (password), something you have (phone), something you are (biometrics). SkaLean supports TOTP (Google Authenticator) and WebAuthn (physical security key).

A mechanism allowing a user to authenticate once to access multiple applications. SkaLean supports SSO via SAML 2.0 and OpenID Connect, enabling integration with Active Directory, Okta, Azure AD and other identity providers.

A structured sequence of automated steps executed to accomplish a business process. In SkaLean AI Automation, a workflow is defined visually with nodes (steps) and connections (data flows). Can include conditions, loops and branches.

An event that initiates the execution of an automated workflow. Examples of SkaLean triggers: email received, form submitted, scheduled time (cron), incoming webhook, CRM event, WhatsApp message received. A workflow can have multiple triggers.

An individual step in an automation workflow. Each node performs a specific action: extract data, call an API, apply a condition, transform, send an email, update a CRM. Nodes are connected to form the execution flow.

A pre-built module allowing SkaLean to connect to a third-party application without development. SkaLean includes 200+ connectors: Salesforce, HubSpot, QuickBooks, Google Workspace, Microsoft 365, Slack, Shopify, Stripe, Twilio, etc.

A system for scheduling task executions at defined times. Cron expressions allow very precise scheduling: "every Monday at 9am", "the 1st of the month", "every 15 minutes". SkaLean supports cron triggers for periodic automations.

A central component that executes automation workflows by coordinating steps, conditions and service calls. SkaLean's AI Automation module includes a native orchestration engine with built-in AI nodes, data sovereignty and turnkey management.

An automatic mechanism that retries the execution of a failed step. SkaLean AI Automation includes native error handling with configurable retry: number of attempts, delay between attempts (exponential backoff), final failure notification.

The AI's ability to identify and extract structured information from unstructured text. Examples: extracting amounts from a PDF invoice, identifying parties in a contract, extracting medications from a prescription. SkaLean uses LLMs for this extraction.

A program simulating a human conversation to answer user questions. SkaLean's AI Agent goes far beyond the traditional chatbot: it accesses your data in real time, executes actions (book, update), and adapts to the context of each conversation.

A customer service approach integrating all communication channels (phone, web chat, email, WhatsApp, SMS, social media) into a consistent experience. SkaLean manages 10 channels from a single control panel with a unified conversation history.

Meta's official API allowing businesses to integrate WhatsApp into their communication systems. SkaLean uses the WhatsApp Business API to deploy AI agents capable of receiving and sending messages on WhatsApp from your professional number.

Technology allowing an AI agent to converse orally via telephone. Combines speech recognition (STT), natural language processing (NLP) and speech synthesis (TTS). SkaLean deploys AI phone agents indistinguishable from a human for appointment booking and support.

Technology converting text into synthetic speech. SkaLean uses high-quality TTS engines for the AI voice of phone agents. Modern voice quality is indistinguishable from a human voice.

Technology converting speech into text. An essential component of SkaLean voice agents. Transcription is done in real time (< 500ms) to allow a fluid conversation. Supports French, Arabic and several regional dialects.

A mechanism by which an AI agent automatically detects situations beyond its capabilities and transfers the conversation to a human agent with the full context. SkaLean configures escalation rules: sensitive topics, detected dissatisfaction, complex requests.

A message sent by the AI agent at the initiative of the system rather than in response to a user. Examples: appointment reminder, order status notification, renewal alert, personalized offer. Requires prior user consent (opt-in).

A metric measuring customer satisfaction after an interaction. Calculated on a 1–5 or 1–10 scale via a simple question. SkaLean automatically collects CSAT after each AI agent interaction and presents trends in the dashboard.

A Reasoning + Acting loop allowing an AI agent to reason step by step before acting. The agent thinks (Thought), decides on an action (Act), observes the result (Observation), then thinks again. Enables complex multi-step reasoning.

Anthropic's open protocol allowing AI agents to connect securely and standardly to external data sources and tools (CRM, calendar, database, APIs). SkaLean implements MCP for real-time agent integrations.

A model for delivering artificial intelligence as a cloud or managed service, avoiding the need for organizations to build and maintain their own AI infrastructure. SkaLean is a sovereign AIaaS: the AI infrastructure is managed by SkaLean, data stays with the client.

Return on investment calculated as (gain - cost) / cost × 100. For SkaLean, typical ROI includes: hours saved × hourly cost + additional revenue generated - SkaLean subscription. Our clients see measurable results from month 2.

Total cost of ownership including all direct and indirect costs of a system over its lifetime: licences, infrastructure, integration, training, maintenance, support. The TCO of a fragmented stack (5 SaaS tools) is generally 3x the listed price.

A software distribution model where the application is hosted in the cloud and accessible via a browser for a subscription fee. ChatGPT, Intercom and HubSpot are SaaS products. SaaS fragmentation (5+ tools) is one of the problems SkaLean solves.

A delivery mode where the provider configures, deploys and trains without the client needing to do any technical work. SkaLean delivers its solutions turnkey: from initial configuration to team training, including integrations with existing tools.

A limited deployment phase of a product with a restricted group of selected clients to validate performance in real conditions. SkaLean is launching its beta pilot in Q2 2026 with 10 selected organizations.

An organization typically counting between 1 and 500 employees, with limited technical and budget resources compared to large enterprises. SkaLean specifically targets SMBs that want AI without hiring data scientists or investing in complex infrastructure.

An appointment or reservation not honoured without prior cancellation. A costly problem in slot-based sectors (medicine, restaurants, hair salons, automotive): every no-show represents lost revenue and a blocked slot. SkaLean automated reminders reduce no-shows by 68%.

A system's ability to handle growing load without performance degradation. For SMBs, AI scalability means being able to increase request volume, automations or channels without hiring proportionally. SkaLean scales automatically according to your needs.

OpenAI's multimodal model (2024) combining text, image and audio understanding. "o" for "omni". Excellent for complex reasoning tasks, document analysis and high-quality content generation. Available in SkaLean.

Anthropic's LLM model renowned for its reliability, balanced responses and resistance to hallucinations. Particularly suited to legal, medical and analytical tasks requiring precision and consistency. Available in SkaLean.

Meta AI's family of open-source LLM models. The key advantage: can be deployed on your own sovereign GPUs. SkaLean uses Llama 3 (8B and 70B parameters) for local sovereign inference, enabling processing without any data transit to external APIs.

A French startup (Paris) developing high-performance open-source LLMs. Mistral 7B and Mixtral 8x7B are recognized for their excellent performance-to-size ratio. Particularly well suited to French. Can be deployed sovereignly on local GPU in SkaLean.

Google DeepMind's family of multimodal LLM models. Available in Flash (fast and economical), Pro (balanced) and Ultra (very powerful) versions. Gemini Pro is available via the Google API in SkaLean for use cases not requiring strict sovereignty.

A mechanism that automatically selects the optimal LLM for each request based on task type, cost constraints, sovereignty and performance. SkaLean implements smart routing: simple requests → fast/economical model, complex requests → powerful model.

A subfield of AI dealing with human natural language. Includes text understanding (NLU), text generation (NLG), translation, sentiment analysis, entity extraction, etc. Modern LLMs are the most advanced form of NLP.

The NLP component enabling a machine to understand the meaning, intent and context of human text. A SkaLean agent uses NLU to understand what the user wants, even when expressed implicitly or with spelling mistakes.

The action or goal a user wants to accomplish, expressed in a message. Example: "I want to cancel my appointment" → intent = cancel_appointment. Intent recognition allows the AI agent to trigger the right response flow.

Specific information extracted from a message: person name, date, number, location, product. In "I want an appointment on March 15 at 2pm", the entities are: date=March 15, time=2pm, action=book_appointment. Entities allow the agent to act with precise data.

Identification of the emotional tone of a text: positive, negative or neutral. SkaLean analyses the sentiment of customer interactions to detect dissatisfaction and automatically trigger escalation to a human agent before the client is lost.

The generation of a concise summary of a long text. SkaLean AI Studio can automatically summarize reports, meeting notes, contracts or emails. Particularly useful for meetings, where the transcript can be summarized with key points and action items.

Translation of text from one language to another by AI. Modern LLMs (GPT-4o, Claude) are excellent translators for common languages. SkaLean can handle multilingual conversations and automatically translate indexed documents.

Technology converting images of text (scanned documents, photos) into usable digital text. SkaLean integrates OCR to index scanned documents, invoice photos or business cards in the AI Studio document base.