The AI terms that
actually matter.
Clear, technical definitions of 113 key concepts — written by AI engineers, not marketers.
GraphRAG
GraphRAG is an advanced retrieval technique that builds a knowledge graph from source documents — extracting entities, relationships, and community summaries — then uses that graph structure to answer complex, multi-hop questions that traditional vector-search RAG cannot handle well.
Read definitionGroup Relative Policy Optimization (GRPO)
GRPO is a reinforcement learning algorithm that trains a policy by sampling a group of outputs per prompt and computing each one's advantage relative to the group's own mean reward, removing PPO's separate critic network entirely. It is the RL backbone behind DeepSeekMath, DeepSeek-R1, and most 2025-2026 open reasoning models.
Read definitionGrouped Query Attention (GQA)
Grouped Query Attention (GQA) is the attention variant behind nearly every serving-optimized LLM today: it splits query heads into groups that each share one key/value head, shrinking the KV cache and memory-bandwidth cost of inference while keeping quality close to full multi-head attention.
Read definitionHalf-Duplex Agent
A half-duplex agent is a voice AI system that speaks and listens one at a time, like a walkie-talkie — waiting for the user to finish before responding — as opposed to a full-duplex agent, which can listen and speak simultaneously.
Read definitionHermes Agent
Hermes Agent is an open-source, self-improving AI agent built by Nous Research that runs as a local gateway process, remembers across sessions, writes its own reusable skills, and reaches you through Telegram, Discord, Slack, WhatsApp, Signal, or the CLI.
Read definitionHNSW (Hierarchical Navigable Small Worlds)
HNSW is a graph-based approximate nearest neighbour (ANN) algorithm that organizes vectors into layered, navigable small-world graphs, enabling logarithmic-time similarity search over billions of embeddings — the default indexing strategy in nearly every production vector database.
Read definitionHuman-in-the-Loop (HITL)
Human-in-the-Loop (HITL) is an architectural pattern and risk management strategy where human judgment, oversight, and intervention are intentionally integrated into an AI system’s lifecycle to ensure accuracy, safety, and alignment.
Read definitionHybrid Search
Hybrid search runs lexical retrieval (typically BM25) and semantic retrieval (dense vector search) against the same query, then fuses the two ranked lists into one, combining exact-term precision with conceptual recall. It has become the default recommendation, not an advanced option, across nearly every major vector database as of 2025-2026.
Read definitionHyDE (Hypothetical Document Embeddings)
HyDE is a zero-shot retrieval technique that uses an LLM to generate a hypothetical document based on a query, and uses that generated document's embedding to find real relevant documents.
Read definitionInflect v2
Inflect v2 is a pair of open-weight, VITS-family English text-to-speech models, Inflect-Micro-v2 (9.36M parameters) and Inflect-Nano-v2 (3.97M parameters), that generate complete 24kHz waveforms locally on CPU with no external vocoder or cloud dependency.
Read definitionJailbreaking
Jailbreaking is the practice of crafting adversarial prompts or multi-turn conversations that manipulate an LLM into bypassing its safety training and producing content or actions it was explicitly aligned to refuse, remaining one of the most persistent unsolved problems in AI safety despite continuous defensive advances.
Read definitionKimi K2.7 Code
A frontier scale, coding focused agentic AI model by Moonshot AI. It features a 1 trillion parameter MoE architecture with advanced multimodality and long horizon software engineering capabilities.
Read definitionStay ahead of the curve
Weekly newsletter on agentic AI, LLMs, and what we're building at Superteams — straight to your inbox.
Ready to ship AI in production?
We deploy fractional AI teams that deliver production-grade systems in 30–90 days. No fluff, no obligation.