Glossary — Azure Agentic Banking AI

Every term used across this track, defined in one line, grouped by area. For depth, follow the linked knowledge module.

Table of Contents

Azure platform

(K00)

  • Azure OpenAI Service — OpenAI models inside Azure's compliance/network/identity boundary; you call a deployment, not a model name.
  • Deployment — a named instance of a model+version+SKU you create and call (e.g. gpt-4o-prod).
  • Azure AI Foundry / Microsoft Foundry — Azure's unified build/eval/deploy/govern platform (2026 rebrand: "Microsoft Foundry"); includes the model catalog and Foundry Agent Service.
  • Foundry Agent Service — managed runtime for production agents, built on the OpenAI Responses API; built-in tools, memory, observability.
  • Azure AI services — the family of prebuilt AI APIs (formerly Cognitive Services).
  • TPM / RPM — tokens-/requests-per-minute quota for a Standard deployment; exceeding → HTTP 429.
  • PTU (Provisioned Throughput Unit) — reserved capacity for predictable latency/throughput, billed monthly.
  • PAYG (Standard) — pay-per-token, shared capacity, variable latency.
  • Entra ID — Microsoft's identity provider (formerly Azure AD).
  • Managed Identity — an auto-managed Entra identity for a resource → keyless auth.
  • RBAC — role-based access control; least-privilege roles on resources.
  • Key Vault — managed secret store.
  • Private Endpoint — gives a PaaS resource a private VNet IP (no public internet).
  • APIM (API Management) — auth/throttle/audit gateway in front of services.
  • Content Safety — harm-category filters + Prompt Shields + Groundedness detection.
  • Region / Data Zone — geographic placement controlling data residency.

GenAI / LLM

(K01)

  • Token — sub-word unit of text (~4 chars EN; ~2× more for Arabic); the unit of cost/latency/context.
  • Embedding — a vector capturing text meaning; similar texts have high cosine similarity.
  • Context window — max tokens (prompt+output) the model attends to at once.
  • Autoregression — generating one token at a time, each conditioned on all prior.
  • Prefill / decode — process the prompt (parallel) / emit tokens (sequential, slow).
  • Temperature — sampling randomness; ~0 = deterministic (the banking default).
  • Function / tool calling — the model requests your code run a function; you execute and feed back.
  • Structured output — constraining the model to JSON matching a schema.
  • Hallucination — a confident, unsupported (possibly false) statement.
  • Fine-tuning — baking style/format into weights (not for changing facts → use RAG).
  • RLHF / DPO — preference alignment training stages.

RAG & retrieval

(K02)

  • RAG — Retrieval-Augmented Generation: retrieve text at query time, generate grounded answer.
  • Azure AI Search — the retrieval engine (formerly Cognitive Search): keyword + vector + hybrid + semantic.
  • Chunking — splitting docs into passages (~300–800 tokens, with overlap).
  • BM25 — keyword ranking (term frequency × inverse document frequency).
  • Vector / ANN / HNSW — semantic search via approximate nearest-neighbor graph.
  • Hybrid search — BM25 + vector combined.
  • RRF (Reciprocal Rank Fusion) — rank-based fusion of result lists (Σ 1/(k+rank)).
  • Semantic ranker — a cross-encoder that re-scores top candidates (retrieve-then-rerank).
  • Integrated vectorization — Azure indexer + embedding skill that chunks/embeds for you.
  • Security trimming — filtering retrieval by the user's access groups (ACLs).
  • Groundedness / faithfulness — every claim supported by retrieved context (the headline RAG metric).

Agents

(K03)

  • Agent — an LLM in a loop that calls tools, observes, and decides next steps.
  • ReAct — reason + act interleaving (think → tool → observe → repeat).
  • LangGraph — stateful graph agent framework (nodes, edges, state, checkpointer, interrupts).
  • LangChain — the building-block toolkit LangGraph is built on.
  • Checkpointer — persists agent state per node → durable, resumable, auditable.
  • Interrupt / human-in-the-loop — pause before a node (e.g. approval), resume later.
  • Semantic Kernel — Microsoft's kernel+plugins SDK (now folded into Agent Framework).
  • AutoGen — Microsoft's multi-agent framework (predecessor to Agent Framework).
  • Microsoft Agent Framework — 2025 convergence of Semantic Kernel + AutoGen; Foundry-native.
  • MCP (Model Context Protocol) — open client–server protocol exposing tools/resources/prompts to any host.
  • Supervisor / multi-agent — specialized agents coordinated by an orchestrator.
  • Orkes — Conductor-based durable workflow/agent orchestration tool.

Documents & vision

(K04, K05)

  • Document Intelligence — Azure IDP service (formerly Form Recognizer): OCR + layout + extraction.
  • OCR / Read — pixels → text + bounding boxes.
  • Layout — text + tables + selection marks + structure (great for RAG chunking).
  • Prebuilt / custom / composed model — ready-made vs trained vs auto-routed extractors.
  • Confidence / STP — per-field score; straight-through-processing rate (auto-accepted %).
  • Bounding region — where a field sits on the page (for highlight/audit).
  • Azure AI Vision — image analysis + Read OCR + Florence models.
  • OpenCV — classic CV library (deskew, threshold, contours, perspective).
  • YOLO — real-time object detection (single forward pass + NMS).
  • Multimodal — model accepting image + text together (GPT-4o vision).
  • RTL / bidi — right-to-left / bidirectional text (Arabic rendering).
  • Eastern Arabic numerals — ٠١٢٣…; normalize to Western before math.
  • MSA vs dialect — Modern Standard vs spoken Gulf/etc. Arabic.

Production & data

(K07)

  • FastAPI — async Python web framework (SSE streaming, Pydantic, OpenAPI).
  • Pydantic — typed validation at the API boundary and for LLM output.
  • SSE — Server-Sent Events; token-by-token streaming.
  • Docker / ACR — container image / Azure Container Registry.
  • Container Apps (ACA) / AKS / ACI — serverless containers / Kubernetes / single-container.
  • Cosmos DB — globally-distributed NoSQL (Document) DB with vector search.
  • Partition key — shards a container; choose high-cardinality, even-access (e.g. conversationId).
  • RU/s (Request Units) — Cosmos throughput/cost currency; exceed → 429.
  • Consistency levels — Strong → Bounded → Session (default) → Consistent Prefix → Eventual.
  • CI/CD — automated build/test/deploy; here with an eval gate.
  • Bicep / IaC — infrastructure as code.
  • OIDC — federated CI auth to Azure without stored secrets.
  • SQL — relational query language (SELECT/JOIN/GROUP BY); see Lab 08.
  • Idempotency key — makes a retried mutating request safe (no double-execute).

Evaluation & safety

(K08)

  • RAGAS — open-source RAG evaluation (faithfulness, relevancy, context precision/recall).
  • Foundry evaluators — Azure-native evaluators (groundedness/safety/agent), offline + continuous.
  • LLM-as-judge — a strong model scoring outputs against criteria (calibrate vs human labels).
  • Recall@k / context precision — retrieval quality metrics.
  • Prompt Shields — Content Safety detection of jailbreak + indirect injection.
  • PII redaction — removing personal data (Presidio / Azure AI Language) before model/logging.
  • Prompt injection — hijacking the model via instructions in user input (direct) or documents (indirect).
  • Red-teaming — adversarial testing (e.g. Microsoft PyRIT).
  • Eval gate — CI step that fails the build on a quality/safety regression.
  • Observability / tracing — App Insights / Azure Monitor spans across prompt→retrieval→tool→model.

Banking & regulation

(K09)

  • KYC / CDD / EDD — Know Your Customer / (Enhanced) Customer Due Diligence.
  • AML / CFT / SAR / PEP — anti-money-laundering / counter-financing-of-terrorism / Suspicious Activity Report / Politically Exposed Person.
  • Core banking system — the system of record for accounts/balances (Finacle/Temenos/Flexcube).
  • IBAN / SWIFT / ISO 20022 / MICR — account number (mod-97 checksum) / wire network / payments standard / cheque line.
  • SR 11-7 — US Fed model-risk-management guidance (dev + independent validation + monitoring).
  • EU AI Act — risk-tiered AI regulation; some financial uses are "high-risk".
  • NIST AI RMF — AI risk management framework.
  • PCI DSS — card-data security standard.
  • Data residency — keeping data in a mandated region/geography.
  • CBUAE / DIFC / ADGM / PDPL — UAE central bank / Dubai & Abu Dhabi financial free zones / UAE data-protection law.
  • Murabaha — a Sharia-compliant (no-interest) financing structure.
  • Human-in-the-loop — human approval/oversight on high-risk or consequential actions.
  • Auditability — being able to reconstruct exactly why the system did something.