H

Hallucination

• Category: Evaluation & safety
• What people say: The model is lying.
• What it actually means: Generated content that is false, unsupported by the available evidence, or inconsistent with the task's source of truth. It can arise even when the output is fluent and the model is not attempting to deceive.
• Why it matters: You usually cannot inspect whether a statement existed in training data, so production checks should focus on support, correctness, and traceability.
• In practice: Require cited evidence for factual answers and evaluate whether each citation actually supports the associated claim.
• Common confusion: A hallucination is an output-quality failure, not a diagnosis of model intent.
• Related terms: Grounding, RAG (Retrieval-Augmented Generation), Verification Gate

Handoff

• Category: AI-native development
• What it actually means: A structured transfer of a task between people or agents that preserves the objective, current state, evidence, decisions, constraints, and remaining work.
• Why it matters: A good handoff prevents the next worker from reconstructing the entire task from a long transcript or repeating completed actions.
• In practice: Pass the accepted plan, changed files, test commands and results, unresolved risks, and exact next action in a compact task packet.
• Common confusion: A summary says what happened. A handoff also says what state is authoritative and what should happen next.
• Learn it: Multi-Session Handoff
• Related terms: Agent State, Checkpoint, Scope Contract, Progressive Disclosure

HNSW

• Category: Retrieval & generation
• Aliases: Hierarchical Navigable Small World
• What it actually means: An approximate-nearest-neighbor index that organizes vectors in layered proximity graphs and searches from coarse upper layers toward detailed lower layers.
• Why it matters: It is a common way to make high-recall vector search practical at scales where exhaustive comparison is too slow.
• In practice: Tune construction and query parameters against latency, memory, and Recall@K targets, then rebuild the index when embedding versions change.
• Common confusion: HNSW is an index algorithm, not a similarity metric, embedding model, or complete vector database.
• Related terms: Approximate Nearest Neighbor (ANN), Vector Database, Embedding, Recall@K
• Sources: Efficient and Robust Approximate Nearest Neighbor Search Using HNSW

Human-in-the-Loop (HITL)

• Category: Agents & tools
• Aliases: Human oversight, human review
• What it actually means: A workflow design in which a person supplies judgment, correction, approval, or escalation at defined points in an AI-driven process.
• Why it matters: Human involvement is most useful at high-impact, ambiguous, or irreversible boundaries, not as an undefined fallback after every step.
• In practice: Let the agent classify routine requests automatically, but route uncertain or high-value cases to a reviewer with the evidence and proposed action.
• Common confusion: HITL does not automatically make a system safe. Reviewers need time, context, authority, and a clear decision standard.
• Related terms: Approval Gate, Verification Gate, Agent, Guardrails

Hybrid Retrieval

• Category: Retrieval & generation
• What it actually means: Retrieval that combines signals from different methods, commonly lexical matching and dense-vector similarity, before merging or reranking results.
• Why it matters: Exact identifiers, rare terms, and semantic paraphrases behave differently, so one retrieval signal can miss useful evidence.
• In practice: Retrieve candidates with both BM25-style keyword search and embeddings, merge their ranks, then rerank the combined set for the user query.
• Common confusion: Hybrid retrieval combines candidate signals. A reranker applies a second relevance model to candidates already retrieved.
• Learn it: Advanced RAG
• Related terms: Semantic Search, Reranker, RAG (Retrieval-Augmented Generation), Embedding

Hyperparameter

• Category: Math & training
• What people say: A setting you tune.
• What it actually means: A configuration choice that shapes model structure, optimization, data processing, or inference rather than being learned as an ordinary model parameter. Examples include learning rate, batch size, layer count, and decoding settings.
• Common confusion: Some hyperparameters are selected before training, while others can be changed during a schedule or at inference time.
• Related terms: Parameter, Learning Rate, Batch Size, Temperature