The AI terms that
actually matter.
Clear, technical definitions of 146 key concepts — written by AI engineers, not marketers.
Query Rewriting
Query rewriting transforms a user's raw query into a different form, a clarified single reformulation, several parallel variants, a higher-level abstraction, or a set of sub-questions, before it reaches the retriever, closing the gap between how people ask and how a corpus is actually indexed. It covers a family of techniques from a small trained rewriter model to LLM-generated multi-query fan-out fused with RRF.
Read definitionQwen 3.5
Qwen 3.5 is Alibaba Cloud’s open-weight LLM series, delivering strong reasoning, long-context support, coding performance, and multilingual capabilities for enterprise AI and agentic workflows.
Read definitionQwen3-ASR
Alibaba's Qwen3-ASR is a state-of-the-art open-source automatic speech recognition model family that outperforms most commercial and open ASR systems on standard benchmarks as of 2026.
Read definitionRadixAttention (SGLang)
RadixAttention is the KV-cache management technique at the core of the SGLang inference engine: it keeps the cached keys and values of every recent prompt and output in a radix tree, so any new request automatically reuses the longest prefix it shares with anything the server has already processed. Paired with a cache-aware scheduler that serves the best-matching requests first, it turns the tree-shaped structure of real LLM traffic (shared system prompts, few-shot examples, multi-turn chats, branching agents) into skipped prefill work.
Read definitionReinforcement Learning from AI Feedback (RLAIF)
RLAIF is a post-training alignment technique that replaces human annotators with a capable AI model to evaluate, critique, and rank outputs, enabling scalable, faster, and more cost-effective alignment of Large Language Models.
Read definitionReinforcement Learning from Human Feedback (RLHF)
RLHF is the training technique that transforms a raw language model into a helpful, instruction-following assistant by using human preference judgements to shape model behaviour — the method behind ChatGPT, Claude, and most frontier chat models.
Read definitionReinforcement Learning with Verifiable Rewards (RLVR)
RLVR is a post-training method that replaces the learned reward model in a standard RLHF pipeline with a deterministic verifier: a program that checks whether an answer is objectively correct. Because the reward comes from running unit tests or matching a known answer rather than from a neural network's opinion, the training signal cannot be gamed the way a learned reward model can, which is why RLVR became the backbone of every open reasoning model released since DeepSeek-R1.
Read definitionReranking
Reranking is a second-stage retrieval step that re-scores and reorders an initial set of candidate documents using a more precise (but more expensive) relevance model, dramatically improving the accuracy of search and RAG pipelines before results reach the LLM.
Read definitionRespect Real-World Ties
A foundational OpenAI safety principle and a measurable training objective directing ChatGPT to actively protect users' real-world relationships, never fostering emotional dependence or replacing human connection.
Read definitionRetrieval-Augmented Generation (RAG)
RAG is technique that combines information retrieval from external sources with text generation, resulting in factually accurate and context-aware AI.
Read definitionRotary Positional Embedding (RoPE)
Rotary Positional Embedding (RoPE) is the positional encoding used by nearly every modern open-weight LLM: it rotates query and key vectors by an angle proportional to their sequence position, so the attention score between two tokens depends only on their relative distance, the property that makes post-training context-length extension (YaRN, NTK-aware scaling, LongRoPE) possible.
Read definitionRRF (Reciprocal Rank Fusion)
Reciprocal Rank Fusion is a simple, training-free algorithm for combining multiple ranked result lists — typically dense vector search and sparse keyword search — into a single, more accurate ranking, using only each item's rank position rather than its raw score.
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.