Glossary

This is the alphabetical reference behind the front-page field index. Each note gives a plain definition, a small example, and the implementation detail an FDE should remember.

A–F

• A2A — discovery and stateful task exchange between independently built agents.
• A2UI — declarative interfaces streamed from agents to trusted renderers.
• Abstention — choosing not to guess when evidence is insufficient.
• Agent — a model-driven loop that can choose actions and tools.
• Agent Card — A2A metadata for an agent’s interfaces, skills, capabilities, and authentication requirements.
• Agent conflict — disagreement between specialist findings with different roles or evidence.
• Agentic RAG — retrieval whose path can change from model decisions and observations.
• Accuracy — the correct share of all classification predictions.
• Answer correctness — whether a response agrees with trusted truth or a reviewed reference.
• Answer relevance — whether a response directly addresses the user’s question.
• BM25 — lexical ranking based on term frequency, rarity, and document length.
• Chain or workflow — a predefined sequence controlled by application code.
• Checkpoint — a persisted snapshot of workflow state.
• Chain-of-thought prompting — using intermediate-reasoning examples for some multi-step tasks.
• Context engineering — curating the instructions, tools, history, state, and evidence a model sees.
• Context window — the token capacity available to one model call.
• Cosine similarity — a normalized vector-direction score, not a relevance probability.
• Embedding — a numeric representation used to compare meaning.
• Faithfulness — whether answer claims are supported by retrieved context.
• Fine-tuning — updating model weights to improve a defined, repeated behavior.
• Feature engineering — turning raw observations into useful model inputs without leaking future data.
• F1 score — a harmonic mean that balances precision and recall for one class.

G–R

• Groundedness — whether answer claims are supported by supplied evidence.
• Graph database — storing entities and explicit relationships for connected-path queries.
• Hallucination — a generated claim unsupported by available evidence.
• HNSW — a graph index for approximate nearest-neighbor search.
• Hybrid search — combining semantic and keyword retrieval.
• LLMOps — evaluation, observability, release, monitoring, and incident practices for LLM systems.
• LoRA — low-rank, parameter-efficient adaptation of model weights.
• Maximal marginal relevance (MMR) — balancing relevance with diversity during result selection.
• Mean reciprocal rank (MRR) — the average reciprocal position of the first relevant result.
• Memory — application-managed information retained and supplied across calls or threads.
• Micro-F1 — F1 after pooling decisions across classes.
• Macro-F1 — the equal-weight mean of per-class F1 scores.
• Message roles — labels separating instructions, user input, model output, and tool results.
• Model Context Protocol (MCP) — a protocol for connecting AI applications to external capabilities.
• Precision@k — the relevant share of the first k retrieved results.
• Prompt — the instructions and context supplied for a task.
• Prompt injection — untrusted content that tries to redirect a model or unsafe action.
• Query rewriting — turning a question into search-friendly wording while preserving intent.
• RLHF — using human preference data as a model-training signal.
• RAG — retrieval followed by evidence-grounded generation.
• ReAct — interleaving model decisions, actions, and observations.
• Recall@k — the known relevant share found in the first k results.
• Reranking — scoring a candidate set more carefully in a second stage.

S–V

• Semantic search — retrieval based on similarity of meaning.
• Semantic cache — reusing an earlier answer for a sufficiently similar query.
• State — data carried from one workflow step to the next.
• System prompt — high-priority instructions for model behavior.
• Thread — an identity grouping one workflow’s saved state.
• Token — a unit of text processed or generated by a model.
• TPM and RPM — token-per-minute and request-per-minute provider limits.
• Tool calling — a structured model request for application code to invoke a capability.
• Vector store — storage and nearest-neighbor retrieval for embeddings and metadata.
• Weighted-F1 — the support-weighted mean of per-class F1 scores.

For longer treatments, continue to production retrieval, agent systems, the applied AI roadmap, or evaluations.