The 212AY Library
AI
Glossary.
200+ artificial intelligence terms explained in plain language — from machine learning and LLMs to AI agents, RAG and prompt engineering. Every definition comes with a concrete example, in English, French and Arabic.
200 terms
Pruning
Pruning is a model compression technique that removes unnecessary weights or neurons from a trained neural network, reducing its size and speed while preserving most of its accuracy. Think of trimming dead branches from a tree to help it grow better. Structured pruning removes entire channels or layers; unstructured pruning zeros out individual weights. Pruning often works alongside quantization to deploy models on resource-constrained devices.
Qdrant
Qdrant is an open-source vector database designed to store embeddings and find the most similar items in milliseconds. It powers semantic search, recommendations, and RAG systems, offering filtering, scaling, and a simple API. Written in Rust for speed, it can run locally, in Docker, or as a managed cloud service. A Casablanca e-commerce shop could use Qdrant to let customers search products by meaning, so a query like «warm winter jacket» returns relevant items even without exact keyword matches.
QLoRA
QLoRA (Quantized Low-Rank Adaptation) is a technique for fine-tuning large language models on modest hardware. It compresses the base model to 4-bit precision through quantization, then trains only small LoRA adapter layers on top, cutting memory needs dramatically while keeping quality close to full fine-tuning. Thanks to QLoRA, a Moroccan startup can adapt an open-source model to Darija customer messages using a single consumer GPU instead of renting an expensive cluster.
Quantization
Quantization reduces the numerical precision of a model's weights, for example from 16-bit to 4-bit numbers, so the model uses less memory and runs faster with only a small loss in quality. It is like compressing a high-resolution photo: the file shrinks but the image stays recognizable. Quantization lets businesses run capable language models on ordinary laptops or local servers. A Moroccan clinic worried about data privacy could run a quantized model on-site instead of sending patient notes to the cloud.
Quantum Machine Learning (QML)
Quantum machine learning explores how quantum computers, which exploit physics phenomena like superposition and entanglement, could speed up or improve machine learning tasks such as optimization and pattern recognition. The field is still experimental: today's quantum hardware is small and noisy, so practical business advantages remain rare. Think of it as research into a future engine, not a tool for tomorrow's project. Banks and logistics firms follow it for long-term problems like portfolio optimization and route planning across African trade networks.
Query Expansion
Query expansion improves search results by enriching the user's original query with synonyms, related terms, or reformulations before retrieval. If a customer types «cheap riad Marrakech», the system may also search for «affordable guesthouse Marrakesh» to catch documents phrased differently. In RAG pipelines, a language model often generates several variants of the question so the retriever finds more relevant passages. It is a low-cost technique that noticeably boosts recall.
Question Answering (QA)
Question answering is the NLP task of automatically producing a direct answer to a question posed in natural language, rather than a list of links. Systems may extract the answer from a given document, retrieve it from a knowledge base, or generate it with a language model grounded in company data. For example, an HR team in Rabat could deploy a QA assistant over its internal policies so employees ask «How many days of paternity leave do I get?» and receive an exact, sourced answer instantly.
Random Forest
A random forest is a machine learning method that combines hundreds of decision trees, each trained on a random slice of the data, then averages their votes to make a prediction. Like polling a crowd of independent experts instead of trusting one, it is more accurate and less prone to overfitting than a single tree. Random forests excel on tabular business data: a Moroccan bank might use one to score credit risk from income, payment history, and employment data.
ReAct Prompting (Reason + Act)
ReAct is a prompting pattern where a language model alternates between reasoning steps («I need the current exchange rate») and actions (calling a search tool or API), then uses the returned observations to reason again. This loop of think, act, observe is the backbone of many AI agents. For example, an agent handling orders for a Casablanca distributor might reason about which warehouse to check, call the inventory API, read the result, and only then confirm delivery.