Artificial Intelligence Glossary 2026

Large-scale language model trained on trillions of tokens, capable of understanding and generating natural language. Examples in 2026: Claude 4.6 (Anthropic), GPT-5 (OpenAI), Gemini 2 (Google), Llama 3+ (Meta), Mistral Large.

Discipline of designing instructions (prompts) to get optimal responses from an LLM. Includes techniques such as role prompting, few-shot learning, chain-of-thought and structured outputs.

💡 SprintMarkt offers a free tool that generates professional prompts by industry and department.

Numeric representation (vector) of text, image or audio that captures its semantic meaning. Two texts with similar meaning will have embeddings that are mathematically close.

Generation by an LLM of false yet seemingly plausible information. Leading cause of poorly implemented AI projects. Mitigated with RAG, validation, guardrails and low temperature.

💡 SprintMarkt applies 3 anti-hallucination layers in chatbots: mandatory RAG, verifiable citations, rule-based post-filtering.

Basic units an LLM processes. Roughly 1 token = 0.75 words in English, 0.5 words in Spanish. API prices are quoted per million tokens.

Maximum number of tokens an LLM can process at once (prompt + response). In 2026 it ranges from 128k (GPT-5) to 1M+ (Gemini Pro, Claude 4.6).

Prompt engineering technique that asks the LLM to 'think step by step' before answering. Dramatically improves mathematical and logical reasoning tasks.

Technique of including 2-5 examples in the prompt so the LLM learns the expected pattern without fine-tuning.

Parameter (0.0-1.0+) that controls the LLM's 'creativity'. 0.0 = deterministic and predictable (ideal for classification). 0.7-0.9 = creative (ideal for generation).

Ability of a model to process and generate multiple input/output types: text, image, audio, video. GPT-4o, Gemini and Claude 4.6 are multimodal.

Technique for generating synthetic speech that imitates a specific voice from just a few seconds of sample audio. Tools: ElevenLabs, Resemble AI, OpenAI Voice.

Family of LLMs developed by Anthropic. In 2026 the active versions are Claude 4.6 (Sonnet, Haiku, Opus). Notable for long-form reasoning, hallucination control and advanced tool use.

Family of LLMs from OpenAI. Active in 2026: GPT-5 (flagship), GPT-4o (vision/audio), GPT-4o-mini (budget). The foundation of ChatGPT.

Family of multimodal LLMs from Google DeepMind. Competitive advantage: native integration with Google Workspace, Search, and the Android ecosystem.

Family of open-weights LLMs from Meta (Llama 3, Llama 4). Permissive license for commercial use. Preferred for self-hosting and scenarios where data cannot leave the client's network.

European LLMs developed in Paris. Models: Mistral Large 2, Mistral Small, Codestral (for code). Advantage: native GDPR compliance, low latency in Europe.

Family of generative image models developed by Stability AI. Open-source (unlike DALL-E or Midjourney), runnable on your own server.

Technique that connects a language model (LLM) to a company's own knowledge base through vector search. The LLM retrieves relevant documents before generating its answer, reducing hallucinations and enabling citations of company-specific sources.

💡 SprintMarkt implements RAG chatbots from €1,125 with a Claude/GPT + pgvector + PostgreSQL stack.

Process of continuing the training of a base LLM with specific data to specialize it in a domain. Unlike RAG (which adds context at inference time), fine-tuning modifies the model's weights.

💡 SprintMarkt recommends RAG over fine-tuning in 90% of SMB cases: cheaper, more flexible, updatable without re-training.

Database specialized in storing and searching numeric vectors (embeddings) with semantic similarity search. 2026 alternatives: pgvector (PostgreSQL), Pinecone, Weaviate, Qdrant, Chroma, Milvus.

💡 SprintMarkt uses pgvector by default in RAG implementations — zero additional cost if the client already runs PostgreSQL.

Ability of an LLM to invoke external functions (APIs, databases, code) during its reasoning. The foundation of 'AI agents'.

Autonomous LLM-based system that receives goals in natural language, plans steps, uses external tools and executes tasks to completion. More complex than a chatbot.

💡 SprintMarkt AI automations from €3,750 and custom AI agents from €13,500 (multi-tool with RAG).

Open protocol proposed by Anthropic in 2024 to connect LLMs to data sources and external tools in a standard way. Ecosystem of 'MCP servers' for GitHub, Notion, Slack, etc.

Feature that forces an LLM to return JSON or another structured format according to a schema. Eliminates post-processing with fragile regexes.

Protection layers applied to LLMs to prevent dangerous, off-topic or policy-violating outputs. Include: input filtering, output validation, prompt injection detection.

Critical comparison. RAG is better when information changes frequently, the dataset is large or sources must be cited. Fine-tuning is better for a specific style/tone or when inference cost matters more.

Model specialized in generating embeddings (numeric vectors) from text. Popular in 2026: text-embedding-3-large (OpenAI), voyage-large-2 (Anthropic/Voyage), bge-large (BAAI, open).

Search based on meaning (via embeddings) instead of exact keywords. Finds 'used car' even when the document says 'second-hand vehicle'.

Running an LLM on your own infrastructure (servers, local GPUs) instead of cloud APIs. Advantage: data control, fixed cost, latency. Disadvantage: requires expensive GPUs + ops expertise.

PostgreSQL extension that adds the 'vector' data type and similarity search. Enables implementing RAG without leaving the client's main relational database.

Standard proposed in 2024 for websites to publish an index optimized for Large Language Models. Works like robots.txt but for AI: it describes the organization, services and key links in Markdown.

Meta's official API for businesses that enables automating bulk messages, chatbots and notifications on WhatsApp. Requires Meta approval and a Business Solution Provider (BSP).

💡 SprintMarkt integrates the WhatsApp Business API as part of AI chatbots (from €1,125).

Open-source, self-hostable visual workflow automation platform. Free alternative to Zapier and Make, with the advantage of full control and no execution limits.

💡 SprintMarkt is an n8n partner and recommends it for self-hosted projects; AI automation from €3,750.

Cloud visual automation platform with ~1,500 integrations. Pricing model based on 'operations' (each individual action). Strong at complex flows with branching.

Cloud automation platform with 7,000+ integrations. Simple and powerful but with aggressive pricing (pay per task). Ideal for simple flows with many SaaS connectors.

Limit on calls per minute/hour imposed by LLM APIs to prevent abuse. Measured in requests/minute (RPM), tokens/minute (TPM) and tokens/day (TPD).

Open-source framework for building applications with LLMs. Provides abstractions for prompts, chains (sequences), agents, memory, RAG. Widely used though criticized for complexity.

Framework specialized in RAG and data ingestion for LLMs. Alternative to LangChain with a more opinionated focus on semantic search over documents.

Platform and community for open-source AI models. Hosts 500,000+ models + datasets. De facto leader of the open-source AI ecosystem.

Open-source tool for running LLMs locally with a single command. Compatible with Llama, Mistral, Phi, Gemma, etc.

European regulation (2024) that classifies AI systems by risk and establishes obligations. Fully in force August 2026. Common obligations: transparency, registration, human oversight, technical documentation.

💡 SprintMarkt includes an AI Use Policy + disclaimers in all its AI projects as part of EU AI Act compliance.

European General Data Protection Regulation (2018). Applies when a system processes data of EU residents. Key obligations: legal basis, consent, right to be forgotten, a DPO where required.

Attack where a user tries to override the system instructions with malicious text ('ignore your previous instructions and...'). Top LLM vulnerability according to the OWASP LLM Top 10.

Emerging discipline (2024-2026) of optimizing content and technical structure to appear in answers from generative engines such as ChatGPT, Claude, Perplexity and Google Gemini. Complements traditional SEO.

💡 SprintMarkt is one of the few agencies in Spain offering an SEO + GEO audit (€490).

Conversational application based on an LLM that answers questions in natural language. Different from classic rule-based chatbots in its ability to understand nuance and context.

💡 SprintMarkt AI chatbots from €1,125 with optional WhatsApp Business API integration.

LLMs are billed per token (input + output). In 2026: Claude Haiku ~$0.25/M input, GPT-4o-mini ~$0.15/M, Gemini Flash ~$0.075/M. 'Flagship' models charge 10-20x more.

Company behind GPT and ChatGPT. Provides an API with GPT-5, GPT-4o, DALL-E, Whisper and embeddings models. Enterprise tier available with HIPAA/SOC2 compliance.

Company behind Claude. Founded by ex-OpenAI staff (2021). Focused on safety and constitutional AI. Claude 4.6 Sonnet/Haiku/Opus models.

OpenAI's conversational product launched in November 2022, built on GPT. Triggered the mass generative AI explosion among the general public.

AI search engine that cites web sources in real time. Alternative to Google for complex queries. Uses a combination of GPT, Claude and in-house models.

Real-time search feature inside ChatGPT (launched 2024). Combines GPT capabilities + web retrieval. Competes with Perplexity and Google.