62 terms
Foundation models
LLM, transformer, attention, embedding, RAG, context window, agent, tool use.
Browse this categoryThe Brocode AI Glossary · 240 terms · 5 categories
Every AI term you'll hear in a UAE boardroom — defined in under 60 seconds.
Plain English, a one-sentence GCC example, and a one-line note on why it matters to your business. No Wikipedia walls of text. Arabic-script parity on Arabic-specific entries. Reviewed quarterly by a named principal engineer.
240 terms
- A
- B
- C
- D
- E
- F
- G
- H
- I
- J
- K
- L
- M
- N
- O
- P
- Q
- R
- S
- T
- U
- V
- W
- X
- Y
- Z
- Foundation models62
- Classical ML48
- MLOps & infrastructure54
- Governance & risk40
- Arabic-specific36
How an entry is built
A consistent three-block format on every term.
The shortest accurate answer, in the same shape every time. Definition. Example. Why it matters.
Block 1 · Plain definition
Forty words or fewer. No jargon. No analogies that require a second analogy. A non-technical reader can quote it into a board paper.
Block 2 · UAE / GCC example
One sentence, regionally grounded. A UAE bank, a federal entity, a Saudi retailer, a Qatari hospital — the example is recognisable to a GCC reader within five seconds.
Block 3 · Why it matters to your business
One sentence on the commercial or operational consequence. If the visitor remembers nothing else, this is the line they remember.
240
Terms in the library
5
Categories — foundation to Arabic-specific
Qtr
Reviewed every quarter by named engineers
60 s
Average time to read an entry
Most-read this month
What other readers searched for first.
- 1RAG
- 2AI agent
- 3Fine-tuning
- 4Data residency
- 5Vector database
- 6Khaleeji Arabic
Browse by category
Five tracks across the 240 terms.
48 terms
Classical ML
Supervised, unsupervised, reinforcement, gradient descent, overfitting, ROC.
Browse this category54 terms
MLOps & infrastructure
Training, inference, drift, observability, model registry, vector database.
Browse this category40 terms
Governance & risk
Model risk, bias, fairness, audit trail, data residency, TDRA, CBUAE, FSRA.
Browse this category36 terms
Arabic-specific
Khaleeji, MSA, NER, tashkeel, tatweel — with Arabic-script parity.
Browse this category
Sample entries — read inline
Fifteen terms in the three-block format.
The first chapter of the glossary, rendered openly. Arabic-script parity is shown on Arabic-specific terms.
Foundation
RAG (Retrieval-Augmented Generation)
- Definition
- A pattern where a language model is fed relevant passages from your own corpus at query time, so its answer is grounded in your data rather than its training memory.
- GCC example
- A UAE bank uses RAG to answer customer queries against its product disclosures — the LLM cites the disclosure page rather than inventing a rate.
- Why it matters
- It is the cheapest way to get an LLM to be useful on private data without retraining.
Foundation
Fine-tuning
- Definition
- Updating a model's weights on a small, task-specific dataset so it performs better on that task. Distinct from RAG, which leaves the weights alone.
- GCC example
- A federal entity fine-tunes Falcon-7B on 8,000 Khaleeji intent-classification examples and improves accuracy from 71 % to 88 %.
- Why it matters
- It is the right answer when behaviour, not knowledge, is what you need to change.
MLOps
Vector database
- Definition
- A database optimised for nearest-neighbour search over high-dimensional embeddings, used as the retrieval layer in most RAG systems.
- GCC example
- A telco stores 4 million product-disclosure passages as 768-dim embeddings in pgvector and serves them at p95 18 ms.
- Why it matters
- Choice of vector store affects latency, cost, and where the data lives — three things procurement asks about.
Foundation
Context window
- Definition
- The maximum number of tokens (roughly, sub-words) a language model can consider at one time. Beyond the window, content is forgotten.
- GCC example
- Claude 3.5 has a 200K-token window; GPT-4o has 128K; Jais 13B has 8K. Window size shapes RAG chunking strategy.
- Why it matters
- Window size is the difference between answering a single email and answering a 60-page board pack.
Foundation
AI agent
- Definition
- A program that uses a language model to decide what tool to call next, in a loop, until a goal is met. Distinguished from a chatbot by its ability to take actions.
- GCC example
- A claims agent in a UAE insurer pulls the policy, queries the fraud model, drafts a settlement letter, and routes it for adjuster review.
- Why it matters
- Agents close the loop between language and action — they are how AI moves from advice to work.
Arabic-specificالعربية الخليجية
Khaleeji Arabic
- Definition
- The cluster of Gulf Arabic dialects spoken across the UAE, KSA, Qatar, Bahrain, Kuwait, and Oman. Distinct from Modern Standard Arabic in vocabulary, morphology, and phonology.
- GCC example
- A Sharjah-based call to a contact-centre is in Khaleeji; an MSA-tuned model mis-transcribes 9–14 % more tokens than a Khaleeji-tuned one.
- Why it matters
- Most production AI for GCC customers fails on Khaleeji unless explicitly tuned for it.
Arabic-specificالعربية الفصحى الحديثة
MSA (Modern Standard Arabic)
- Definition
- The pan-Arab literary and broadcast register, used in print, news, and official documents. Almost nobody speaks it in everyday conversation.
- GCC example
- A federal-entity letter is in MSA; the call-centre conversation about that letter is in Khaleeji.
- Why it matters
- Most published Arabic NLP benchmarks score MSA — your production data is rarely MSA.
Arabic-specificالتشكيل
Tashkeel
- Definition
- Diacritical marks that disambiguate Arabic vowels and grammatical roles. Usually absent in everyday text; sometimes essential for disambiguation.
- GCC example
- Without tashkeel, the word "كتب" can mean "he wrote", "books", or "was written" — a Khaleeji intent classifier has to handle all three.
- Why it matters
- Tashkeel handling is a tokenisation question with downstream model-quality effects.
Governance
Data residency
- Definition
- A requirement that data remain physically within a specified jurisdiction at rest, in transit, and during processing.
- GCC example
- A CBUAE-supervised bank requires customer PII to remain in UAE-resident infrastructure — practically, in AWS UAE North, Azure UAE North, OCI Abu Dhabi, or G42 Cloud.
- Why it matters
- Residency drives architecture, vendor selection, and contract clauses. Get it wrong at design and you pay at audit.
Foundation
Hallucination
- Definition
- When a language model produces a fluent, plausible answer that is factually wrong — usually because the answer was not grounded in retrieved evidence.
- GCC example
- An ungrounded chatbot tells a customer a fictitious branch is open at midnight; the customer drives there and complains.
- Why it matters
- Hallucinations are the failure mode procurement most worries about. RAG, citation, and refusal patterns reduce them — not eliminate them.
MLOps
Model drift
- Definition
- A gradual deterioration in model performance because the live data distribution has drifted away from the training data.
- GCC example
- A fraud model trained pre-Ramadan misses a seasonal mule-account pattern; precision falls 12 % in the first week of the holy month.
- Why it matters
- Drift detection is the single MLOps capability separating a hobby model from a production one.
Governance
Model card
- Definition
- A short structured document recording what a model does, what data trained it, how it was evaluated, and its known limits — the regulator-facing equivalent of a datasheet.
- GCC example
- A CBUAE-supervised bank publishes a model card for every AI model in customer-facing production. The card is the first thing the supervisor asks for.
- Why it matters
- Model cards turn AI risk into a documented, auditable artefact. No card, no production deployment.
Foundation
Embedding
- Definition
- A vector of numbers representing the meaning of a piece of text, image, or audio in a high-dimensional space — so that semantically similar items sit near each other.
- GCC example
- A retailer embeds product titles in Arabic and English in a shared 768-dim space; a Khaleeji query finds the English product.
- Why it matters
- Embeddings are how search escapes keyword matching and starts behaving like a human reader.
Governance
TDRA-compliance
- Definition
- Conformance with the technical and licensing standards set by the UAE Telecommunications and Digital Government Regulatory Authority for connected and AI-enabled services.
- GCC example
- A federal-entity AI service exposes APIs through a TDRA-licensed interconnect; data does not leave the UAE.
- Why it matters
- TDRA-alignment is a procurement gate for federal projects. It shapes architecture and operating model.
MLOps
Quantisation
- Definition
- Reducing the numeric precision of a model's weights — for example, from 16-bit to 8-bit or 4-bit — to shrink memory and accelerate inference.
- GCC example
- A 13B-parameter LLM quantised to INT4 fits on a single 24 GB GPU and serves at twice the throughput.
- Why it matters
- Quantisation is how a self-hosted model becomes commercially viable on a sensible GPU bill.
Reviewed by Yasmin Al Marzooqi, Head of Arabic NLP — last refresh February 2026.
Arabic-specific entries — English and Arabic side-by-side
Five regional terms in script-parity format.
Where the term is anchored in the regional context, the Arabic rendering sits beside the English. Reviewed by a native MSA editor.
Khaleeji Arabic
العربية الخليجية
The cluster of Gulf Arabic dialects spoken across the UAE, KSA, Qatar, Bahrain, Kuwait, and Oman. Distinct from Modern Standard Arabic in vocabulary, morphology, and phonology.
Why it matters: Most production AI for GCC customers fails on Khaleeji unless explicitly tuned for it.
MSA (Modern Standard Arabic)
العربية الفصحى الحديثة
The pan-Arab literary and broadcast register, used in print, news, and official documents. Almost nobody speaks it in everyday conversation.
Why it matters: Most published Arabic NLP benchmarks score MSA — your production data is rarely MSA.
Tashkeel
التشكيل
Diacritical marks that disambiguate Arabic vowels and grammatical roles. Usually absent in everyday text; sometimes essential for disambiguation.
Why it matters: Tashkeel handling is a tokenisation question with downstream model-quality effects.
Free download
Brocode AI Glossary — Pocket Guide
A 24-page downloadable distillation of the 60 most-asked terms, formatted for printing or reading on a phone. The three-block format is preserved. The Arabic-script parity is preserved.
- The 60 most-asked terms in the three-block format
- Foundation, Classical ML, MLOps, Governance, Arabic-specific
- Arabic-script parity on Arabic-specific terms
- Printable single-sheet quick-reference at the back
- Reviewed by Yasmin Al Marzooqi — last refresh February 2026
Monthly vocabulary email
Three new terms. One practitioner guide. One minute to read.
One short email per month. You can unsubscribe in one click, and we will not put you on a sales rotation.
Glossary editorial
Missing a term? Suggest one — the editorial team reviews suggestions monthly.
Common requests in the last quarter: speculative decoding, ICV (in-country value), tatweel, prompt injection, and federated learning. Three of these will be in the next refresh.
Suggest a termNext steps
Related capabilities and stories
Insights hub
Whitepapers, guides, FAQs, blog, case studies — the wider library.
Read morePractitioner guides
Each glossary term links to the guide that goes deeper.
Read moreCommercial FAQs
When the question moves from "what is this?" to "how does Brocode answer this?".
Read moreNLP services
For Arabic-specific terms with a commercial reflection.
Read moreGenerative AI development
For LLM-related terms with a commercial reflection.
Read more