8. Comprehensive Glossary

Cross-references point to the subsection where each term is developed.

• A2A (Agent-to-Agent) — Standard for communication between agents; complementary to MCP. (2.3.2)
• Agent — A model plus instructions, tools, knowledge, and a persistent thread; can take multi-step actions, unlike a bare completion. (1.1.2, 2.2.1)
• Agent Service (Foundry) — Managed runtime hosting agents, handling state, identity, and observability. (2.2)
• BM25 — Lexical keyword-ranking algorithm; the keyword half of hybrid search. (4.2.1)
• Chunking — Splitting documents into passages for embedding/retrieval; fixed-size (Text Split) or structure-aware (Document Layout). (4.1.2)
• Code Interpreter — Hosted sandboxed Python tool for data analysis, math, and charts; not a gateway to your systems. (2.2.3)
• Coherence (evaluator) — Measures whether output is well-formed language. (2.5.1)
• Connected agent — An agent invoked by another agent as a building block of orchestration. (2.3.1)
• Content Safety / harm categories — Filtering across hate, sexual, violence, self-harm at configurable severities. (3.3.2)
• Content Understanding — Multimodal extraction across documents, images, audio, video. (4.3.1)
• Cross-encoder — Processes query and document together (used by semantic ranker); higher nuance than a bi-encoder. (4.2.1)
• Customer-managed keys (CMK/BYOK) — Encryption keys you control/audit in Key Vault; needed only when key control is required. (3.2.2)
• DefaultAzureCredential — Credential chain enabling secret-free auth via managed identity in production. (2.1.1, 3.2.1)
• Deployment — A specific model+version with quota, deployed into a project; you call its endpoint. (1.2.2)
• Deployment type — Capacity model (Standard / Provisioned / Batch) × routing scope (Global / Data Zone / Regional). (3.1.1)
• Document Intelligence — Service extracting structure/fields from documents (layout, prebuilt, custom). (4.3.1)
• Embedding — Vector representation of text capturing meaning; query and index must use the same model. (4.1.2)
• Fine-tuning — Adjusts a model's learned behavior/format; never adds facts (use grounding for facts). (1.3.2, 2.5.2)
• Foundry (Microsoft) — Unified platform: model catalog, Agent Service, evaluations, connections, governance. (1.1.1, 1.2)
• Function (tool) calling — Model emits a structured request to run a function; your app executes it. (2.1.3, 2.2.3)
• GenAIOps — DevOps for generative apps plus evaluation gates and AI observability. (3.4.2)
• Grounding — Supplying retrieved/attached data at request time so the model reasons over real facts. (1.3, 4.1.1)
• Groundedness (evaluator/detection) — Checks whether output is supported by the source; catches hallucination. (2.5.1, 3.3.2)
• Hybrid search — Keyword + vector results fused via Reciprocal Rank Fusion for maximum recall. (4.2.1)
• Integrated vectorization — Built-in chunk+embed in the Search indexer pipeline at index and query time. (4.2.2)
• Knowledge source — Data an agent grounds on (AI Search index, files, Bing). (2.2.1)
• Managed identity — Azure-managed credential with no stored secret; preferred over API keys. (3.2.1)
• MCP (Model Context Protocol) — Standard for connecting an agent to tools/data; complementary to A2A. (2.3.2)
• Multimodal model — Model accepting image/other inputs alongside text; handles most visual understanding. (5.1.1)
• Prompt Shields — Defense against direct jailbreaks and indirect prompt injection. (3.3.2)
• Provisioned throughput (PTU) — Reserved, model-independent capacity billed hourly; for predictable load. (3.1.1, 3.1.2)
• Quota vs. capacity — Quota is permission to deploy; capacity is allocated at deployment time and not guaranteed by quota. (3.1.2)
• RAG — Retrieval-Augmented Generation: retrieve relevant chunks, ground the model on them. (4.1)
• RBAC (least privilege) — Grant the narrowest role for the task (e.g., Cognitive Services OpenAI User). (3.2.1)
• Read (OCR) — Purpose-built printed/handwritten text extraction with per-word confidence. (5.1.2)
• Relevance (evaluator) — Measures whether output addresses the query. (2.5.1)
• Run — One execution of an agent against a thread; may include multiple tool-call steps. (2.2.2)
• Semantic ranking — Cross-encoder re-scoring of top results for precision/nuance. (4.2.1)
• Structured output (JSON mode/schema) — Constrains response shape; schema-valid is not fact-valid. (2.1.3)
• Temperature — Controls randomness/diversity, not capability; low for factual tasks. (2.1.2)
• Thread — Persistent conversation/state container an agent runs against. (2.2.2)
• TPM quota — Tokens-per-minute ceiling set per standard deployment. (3.1.2, 3.3.1)
• Tracing (OpenTelemetry) — Step-level spans across a run; explains agent behavior metrics can't. (3.4.1)
• Vector search — Bi-encoder semantic similarity matching meaning, not exact terms. (4.2.1)