AI Prompting Glossary

A prompting technique where the AI model is asked to perform a task without being provided any prior examples. The model relies entirely on its pre-trained knowledge to generate the response.

A technique where the AI is provided with a small number of examples (usually 2 to 5) within the prompt to demonstrate the desired format, style, or logic before asking it to perform the actual task.

A prompting strategy that forces the AI to break down complex, multi-step problems into intermediate reasoning steps before arriving at a final answer.

A phenomenon where an AI model generates false, nonsensical, or unverified information, presenting it confidently as fact. Hallucinations are often mitigated by providing strong context (RAG) or using structured frameworks like STCO.

An architecture that connects an LLM to an external knowledge base or database. Instead of relying on its training data, the model retrieves the exact factual documents needed to answer a query.

A parameter in AI inference that controls the randomness of the output. A temperature of 0 produces deterministic, highly predictable responses (good for coding/legal), while a temperature of 1 produces highly creative and varied responses (good for brainstorming/poetry).

System, Task, Context, Output. A proprietary prompt engineering framework ensuring perfectly aligned AI generation by defining exactly who the AI is (System), what it must do (Task), what data constraints it has (Context), and the exact format it should return (Output).

A technique where the output of one prompt is used as the input to the next, creating a sequential pipeline of AI operations. Each step handles a discrete sub-task, improving accuracy on complex workflows that would overwhelm a single prompt.

An advanced reasoning framework that extends Chain of Thought by exploring multiple reasoning paths simultaneously, evaluating each branch, and selecting the most promising one. It mimics deliberate planning rather than linear, left-to-right generation.

A prompting paradigm that interleaves reasoning traces with concrete actions (such as searching a database or calling an API). The model thinks about what to do, executes the action, observes the result, and then reasons again — enabling tool-augmented problem solving.

A decoding strategy where the model generates multiple independent answers to the same prompt using different reasoning paths, then selects the most frequent or consistent answer. This significantly improves reliability on mathematical and logical tasks.

A technique where the prompt explicitly assigns a persona, expertise level, or professional identity to the AI model. This steers the model's tone, vocabulary, depth, and perspective to match a specific domain expert.

A higher-order technique where a prompt instructs the AI to generate, refine, or optimise other prompts. The model acts as a prompt engineer itself, iterating on instructions to improve downstream task performance.

A technique, commonly used in image and text generation, where the prompt explicitly specifies what the output should NOT contain. This constrains the model away from undesirable outputs, hallucinations, or stylistic choices.

A fine-tuning methodology where a pre-trained language model is further trained on a dataset of (instruction, response) pairs. This aligns the model to follow human instructions more accurately and is the technique behind models like InstructGPT and Gemini.

The process of taking a pre-trained foundation model and training it further on a smaller, domain-specific dataset to specialise its performance. Fine-tuning adjusts the model's weights to improve accuracy on targeted tasks without training from scratch.

A training technique where human evaluators rank model outputs by quality, and a reward model is trained on those rankings. The language model is then optimised via reinforcement learning to produce outputs that align with human preferences for helpfulness, safety, and accuracy.

Dense numerical vector representations of text, images, or other data that capture semantic meaning in a high-dimensional space. Semantically similar items have vectors that are close together, enabling similarity search, clustering, and retrieval-augmented generation.

A specialised database optimised for storing, indexing, and querying high-dimensional embedding vectors at scale. Vector databases power semantic search and RAG pipelines by enabling fast approximate nearest-neighbour lookups.

The process of converting raw text into discrete units (tokens) that a language model can process. Tokens may represent whole words, subwords, or individual characters depending on the tokenizer. Token count directly affects cost, latency, and context window usage.

The maximum number of tokens a language model can process in a single request, encompassing both the input prompt and the generated output. Larger context windows allow processing longer documents but increase computational cost.

The process of running a trained model to generate predictions or outputs from new input data. In the context of LLMs, inference is the act of generating text token-by-token in response to a prompt. Inference cost and latency are key production considerations.

The technique of connecting a language model's generation to verifiable, external sources of truth — such as search results, databases, or uploaded documents — to reduce hallucinations and improve factual accuracy.

A decoding parameter that restricts token selection to the smallest set of tokens whose cumulative probability exceeds the threshold P. Unlike Top-K, it dynamically adjusts the candidate pool size based on the probability distribution, balancing creativity and coherence.

A decoding parameter that limits the model to sampling from only the K most probable next tokens at each generation step. Lower values produce more focused, predictable output; higher values introduce more variety.

A parameter that sets the upper limit on the number of tokens the model will generate in its response. It controls output length and directly affects API cost. The parameter does not guarantee the model will use all tokens — it sets a ceiling.

Specific strings or tokens that, when generated by the model, signal it to immediately stop producing further output. Stop sequences give precise control over where the model's response ends, preventing runaway generation.

A parameter that penalises tokens proportionally to how often they have already appeared in the generated text. Higher values discourage repetition, making the model less likely to reuse the same words and phrases.

A parameter that applies a flat penalty to any token that has already appeared in the output, regardless of how many times. Unlike frequency penalty, it does not scale with repetition count — it simply encourages the model to introduce new topics and vocabulary.

AI systems that can autonomously plan, execute multi-step tasks, use tools, and make decisions with minimal human intervention. Agentic architectures combine LLMs with memory, tool access, and goal-directed reasoning to complete complex workflows end-to-end.

AI models capable of processing and generating multiple data types — including text, images, audio, video, and code — within a single unified architecture. Multi-modal models understand cross-modal relationships, such as describing an image or generating code from a sketch.

The foundational neural network architecture behind modern LLMs, introduced in the 2017 paper 'Attention Is All You Need.' Transformers process entire sequences in parallel using self-attention mechanisms, replacing recurrent networks and enabling massive scaling.

A core component of the Transformer architecture that allows the model to weigh the relevance of every token in the input when generating each output token. Attention enables the model to focus on the most contextually important parts of long inputs.

Capabilities that appear in large language models only above certain scale thresholds and are not present in smaller models. These abilities — such as multi-step reasoning, code generation, and in-context learning — emerge unpredictably as model size and training data increase.

A phenomenon in machine learning where a model trained on new data loses performance on previously learned tasks. This is a critical challenge in fine-tuning and continual learning, where specialising the model on new domains can degrade its general capabilities.

A security vulnerability where malicious instructions are embedded within user input or external data to override the model's system prompt and intended behaviour. Prompt injection can cause the model to leak confidential instructions, bypass safety filters, or execute unintended actions.

The practice of crafting prompts specifically designed to bypass an AI model's safety guardrails, content policies, or usage restrictions. Jailbreaking exploits weaknesses in alignment training to make the model produce outputs it was explicitly trained to refuse.

A structured adversarial testing practice where human evaluators or automated systems deliberately attempt to provoke harmful, biased, or policy-violating outputs from an AI model. Red teaming uncovers vulnerabilities in safety alignment, prompt injection defences, and content filters before deployment.

Programmatic or model-level constraints that enforce safety, compliance, and quality boundaries on AI outputs. Guardrails can be implemented as system prompt rules, output validators, content classifiers, or dedicated filtering layers that intercept and block harmful or off-topic responses.

The discipline of designing and optimising the entire information payload delivered to an LLM — including system prompts, retrieved documents, conversation history, tool outputs, and structured metadata — to maximise response quality within the model's context window. Context engineering goes beyond prompt engineering by treating the full input as a designed system.

The fundamental unit of text that a language model processes. A token can be a whole word, a subword fragment, a single character, or a special symbol. Token count determines API cost, latency, and how much content fits within a model's context window.

A simplified alternative to RLHF that trains a language model to align with human preferences without requiring a separate reward model. DPO directly optimises the policy model using a binary cross-entropy loss on preference pairs, making alignment training more stable and computationally efficient.

A parameter-efficient fine-tuning technique that freezes the pre-trained model weights and injects small, trainable low-rank matrices into each transformer layer. LoRA dramatically reduces the number of trainable parameters (often by 99%+), enabling fine-tuning of large models on consumer hardware.

An extension of LoRA that combines 4-bit quantization of the base model with low-rank adapters. QLoRA enables fine-tuning of models with billions of parameters on a single consumer GPU by dramatically reducing memory requirements while maintaining comparable quality to full fine-tuning.

A model compression technique where a smaller 'student' model is trained to replicate the behaviour and outputs of a larger 'teacher' model. The student learns from the teacher's soft probability distributions rather than just hard labels, capturing nuanced patterns that transfer knowledge efficiently into a compact model.

A neural network architecture that uses multiple specialised sub-networks (experts) and a gating mechanism that routes each input to only the most relevant experts. MoE models achieve the quality of very large dense models while using only a fraction of the parameters during inference, dramatically reducing compute cost.

A search technique that retrieves results based on the meaning and intent behind a query rather than exact keyword matching. Semantic search uses embedding vectors to compare the conceptual similarity between queries and documents, enabling more accurate and contextually relevant retrieval.

A compressed, abstract mathematical representation of data learned by a neural network. In latent space, similar inputs cluster together and meaningful operations (like interpolation between concepts) become possible. LLMs encode the entire meaning of an input sequence into positions within this high-dimensional space.

A prompting technique that combines multiple input types — such as text, images, audio, video, or code — within a single prompt to leverage a multimodal model's cross-modal understanding. This enables tasks that require interpreting visual context alongside textual instructions.

A capability where an LLM can recognise when it needs to invoke external tools — such as APIs, calculators, databases, or code interpreters — and generates structured function calls with the correct arguments. The model reasons about which tool to use, calls it, receives the result, and incorporates it into its response.

A reusable prompt structure with placeholder variables that get filled in dynamically at runtime. Templates standardise AI interactions across an application, ensuring consistent quality and format while allowing customisation per request.

Techniques for reducing the token count of a prompt without losing essential information, enabling more content to fit within the model's context window and reducing API costs. Methods include summarisation, selective extraction, and learned compression models like LLMLingua.

A class of AI model that jointly processes and reasons over both visual (image/video) and textual inputs within a single architecture. VLMs can describe images, answer questions about visual content, extract text from documents, and generate responses grounded in visual context.

An alignment technique developed by Anthropic where the model is trained to self-critique and revise its own outputs based on a set of written principles (a 'constitution'). Instead of relying solely on human feedback, the model learns to evaluate whether its responses align with explicit ethical, safety, and quality guidelines.

An AI system that independently plans, executes, and iterates on complex multi-step tasks by combining a language model with persistent memory, tool access, and goal-oriented reasoning loops. Unlike simple chatbots, autonomous agents can operate over extended periods, manage sub-tasks, and recover from errors without human guidance.

A standardised evaluation dataset and methodology used to measure and compare AI model performance on specific tasks. Benchmarks provide objective, reproducible scores that enable researchers and practitioners to assess progress, compare models, and identify strengths and weaknesses.

A privileged instruction block set by the application developer (not the end user) that defines the AI model's persona, behaviour rules, output format, and safety constraints. The system prompt persists across the entire conversation and takes priority over user messages in most models.

The message or instruction provided by the end user in a conversation with an AI model. The user prompt represents the actual query, task, or request that the model should respond to, within the constraints established by the system prompt.

The AI model's response in a conversation, which can also be pre-filled by the developer to steer the model's output format, tone, or starting direction. Pre-filling the assistant prompt is a technique for controlling output structure without adding to the system prompt.

The practice of optimising web content to be cited, quoted, or surfaced by AI-powered answer engines — such as ChatGPT, Perplexity, Google AI Overviews, and Gemini — rather than traditional search engine results pages. AEO focuses on structured data, clear definitions, FAQ markup, and authoritative sourcing to maximise AI citation probability.

Machine-readable markup added to web pages that explicitly describes the content's meaning to search engines and AI systems. Structured data uses standardised vocabularies (primarily Schema.org) to label entities, relationships, and attributes, enabling rich results, knowledge graph integration, and AI citation.

The recommended format for embedding structured data in web pages. JSON-LD uses a script tag to inject machine-readable metadata without altering the visible HTML content, making it the preferred method by Google and AI answer engines for consuming structured information.

The total number of tokens a model can accept in a single request, encompassing system prompt, conversation history, retrieved documents, and the expected output. Context length is a hard architectural limit determined during training and directly constrains how much information can influence a single generation.

Strategies for efficiently utilising a model's finite context window by prioritising the most relevant information, compressing less critical content, and structuring the payload to maximise response quality. Effective management prevents context overflow and ensures the model attends to the right information.

The ability of large language models to learn new tasks at inference time purely from examples and instructions provided within the prompt, without any weight updates or fine-tuning. ICL is the mechanism that makes few-shot and zero-shot prompting work.

The operational discipline of managing prompts in production systems, encompassing version control, A/B testing, monitoring, evaluation pipelines, and deployment workflows. PromptOps applies DevOps and MLOps principles to prompt engineering, treating prompts as critical production assets.

A development practice where prompts are stored, versioned, and managed as code artefacts within a repository rather than being hard-coded strings in application logic. This enables version control, code review, automated testing, and environment-specific prompt configurations.

The practice of maintaining a history of prompt changes with semantic versioning, enabling teams to track which prompt version produced which outputs, roll back to previous versions, and A/B test prompt variations in production.

A centralised repository or service that stores, catalogues, and serves prompt templates across an organisation. A prompt registry provides a single source of truth for all production prompts, enabling discovery, reuse, and governance at scale.

A framework for programming — rather than manually prompting — language models. DSPy replaces hand-crafted prompt strings with declarative Python modules that are automatically optimised through compilation, enabling systematic prompt optimisation backed by metrics rather than trial-and-error.

An evaluation paradigm where a language model is used to assess the quality, correctness, or safety of another model's outputs. The judge model scores responses against defined criteria, providing scalable evaluation that approximates human judgment at a fraction of the cost.

The systematic process of measuring a language model's performance across dimensions such as accuracy, relevance, fluency, safety, and latency. Evaluation combines automated metrics (BLEU, ROUGE, exact match), model-based judging, and human assessment to determine fitness for production use.

An architecture pattern that dynamically directs incoming requests to different language models based on task complexity, cost constraints, latency requirements, or content type. Routing optimises the cost-quality trade-off by using cheaper models for simple tasks and more capable models for complex ones.

A structured approach to prompt security that encompasses System hardening, Harmful input detection, Input validation, Encapsulation of sensitive data, Layered defence strategies, and Detection monitoring. SHIELD provides a comprehensive defence-in-depth model for protecting LLM-powered applications from adversarial attacks.

A prompt injection defence pattern where critical system instructions are placed both before and after the user's input, 'sandwiching' potentially malicious content between authoritative instructions. This makes it harder for injected instructions to override the system prompt because the model encounters reinforcing instructions after processing user input.

The practice of programmatically scanning, filtering, and sanitising user inputs before they reach the language model. Input validation detects and blocks prompt injection attempts, removes harmful patterns, enforces length limits, and strips special characters that could manipulate model behaviour.

A collection of defensive design patterns used to protect LLM applications from adversarial exploitation, including prompt injection, data exfiltration, and jailbreaking. These patterns encompass input sanitisation, output filtering, privilege separation, instruction hierarchy, and monitoring.

A prompting technique where exactly one example is provided to demonstrate the desired task format before presenting the actual query. One-shot sits between zero-shot (no examples) and few-shot (multiple examples), offering a minimal demonstration when token budget is limited.

An extension of few-shot prompting that leverages large context windows to include dozens or even hundreds of examples in a single prompt. Many-shot prompting can match or exceed fine-tuning performance on specific tasks by providing extensive in-context demonstrations.

A technique that generates multiple responses using different prompt variations, phrasings, or model configurations, then combines them to produce a more robust final answer. Ensembling reduces variance and improves reliability by aggregating diverse perspectives on the same task.

The systematic process of diagnosing why a prompt produces incorrect, incomplete, or inconsistent outputs. Debugging techniques include isolating prompt components, testing with controlled inputs, analysing attention patterns, inserting reasoning checkpoints, and comparing outputs across model versions.

The process of extracting structured data from a language model's free-text response. Output parsing transforms raw model output into programmatically usable formats (JSON, objects, arrays) using regex, schema validation, or specialised parsing libraries like LangChain's output parsers.

A technique where learnable continuous vectors (soft tokens) are prepended to the input and optimised through gradient descent, rather than using discrete natural-language instructions. Soft prompts are task-specific and often outperform hand-written prompts while being much smaller than full fine-tuning.

The standard practice of writing discrete, natural-language instructions as prompts for a language model. Hard prompts are human-readable, interpretable, and manually crafted — in contrast to soft prompts, which are continuous learned vectors not directly interpretable by humans.

A prompting pattern where relevant documents or data are dynamically retrieved from external sources and injected into the prompt context before generation. Unlike full RAG architectures, this refers specifically to the prompt construction technique of enriching queries with retrieved context.

A retrieval technique where the original user query is automatically augmented with synonyms, related terms, or LLM-generated reformulations to improve recall in search and RAG pipelines. Query expansion ensures relevant documents are retrieved even when the user's phrasing differs from the indexed content.

A retrieval strategy that combines traditional keyword-based search (BM25/TF-IDF) with semantic vector search (embeddings) to leverage the strengths of both approaches. Hybrid search improves retrieval quality by catching both exact-match terms and semantically related content.

A specialised neural network that converts text, images, or other data into dense numerical vectors (embeddings) in a high-dimensional space. Unlike generative LLMs, embedding models are optimised specifically for producing representations where semantic similarity corresponds to vector proximity.

A quantitative measure of how semantically related two pieces of content are, calculated by comparing their embedding vectors using distance metrics such as cosine similarity, dot product, or Euclidean distance. Higher similarity scores indicate closer semantic meaning.

The process of fine-tuning a pre-trained language model on a labelled dataset of (input, desired output) pairs using standard supervised learning. SFT is typically the first alignment step after pre-training and before RLHF, teaching the model to follow instructions and produce helpful responses.

A family of fine-tuning techniques that update only a small subset of a model's parameters while freezing the rest, dramatically reducing compute and memory requirements. PEFT methods include LoRA, QLoRA, prefix tuning, and adapter tuning — enabling customisation of large models on limited hardware.

A parameter-efficient fine-tuning method that inserts small trainable neural network modules (adapters) between the frozen layers of a pre-trained model. Each adapter typically contains a down-projection, non-linearity, and up-projection, adding less than 5% additional parameters.

A parameter-efficient fine-tuning technique that prepends a sequence of trainable continuous vectors (prefixes) to the key and value matrices at every transformer layer. Only the prefix parameters are trained, while the entire pre-trained model remains frozen.

An alignment method that combines supervised fine-tuning and preference alignment into a single training stage by incorporating an odds-ratio-based penalty into the SFT loss. ORPO eliminates the need for a separate reference model required by DPO, simplifying the training pipeline and reducing compute costs.

A preference optimisation method inspired by prospect theory that aligns language models using only binary feedback (good/bad) rather than requiring paired comparisons. KTO is more practical than DPO in real-world settings where collecting pairwise preferences is expensive, but simple thumbs-up/thumbs-down signals are readily available.

The practice of crafting effective prompts specifically for image understanding tasks with vision-language models. Vision prompting techniques include providing spatial references, requesting structured visual analysis, specifying inspection regions, and combining text instructions with visual markers or annotations.

Techniques for instructing multimodal models to process, transcribe, analyse, or generate audio content. Audio prompting involves specifying the desired audio analysis task, output format, and any domain-specific considerations such as speaker diarisation, language detection, or emotional tone analysis.

An open protocol that standardises how AI applications provide context to language models by defining a universal interface for connecting models to external data sources, tools, and services. MCP eliminates custom integrations by providing a consistent API that any compliant tool or data source can implement.

A technique where the computed key-value representations of frequently reused prompt prefixes (such as system prompts or static context) are stored and reused across requests, avoiding redundant computation. Prompt caching significantly reduces latency and cost for applications with shared prompt components.

An inference optimisation that stores the computed key and value tensors from the attention mechanism for all previously processed tokens. The KV cache allows the model to generate each new token by computing attention only against the new token and the cached representations, rather than reprocessing the entire sequence.

A model capability or API feature that constrains the language model to generate output conforming to a specified schema, such as JSON, XML, or a custom format. Structured output guarantees parseable, schema-valid responses, eliminating the need for fragile regex-based output parsing.

An API setting that constrains the model to output only valid JSON in its response. JSON mode ensures the output is always parseable but does not enforce a specific schema — for schema enforcement, use structured output with a JSON Schema definition.

A class of language model specifically trained to perform extended, multi-step logical reasoning before producing a final answer. Reasoning models (like OpenAI's o1 and o3, or Google's Gemini 2.5 Pro) use internal chain-of-thought processes that significantly improve performance on mathematics, science, coding, and complex analysis tasks.

A capability in reasoning models where the model allocates additional compute to 'think' through complex problems before responding. The thinking process generates internal reasoning tokens that are not shown to the user but significantly improve answer quality on hard tasks. Users can often configure a thinking budget.

The system by which an AI agent stores, retrieves, and utilises information across interactions and tasks. Agent memory encompasses short-term working memory (current context), episodic memory (past interactions), semantic memory (learned facts), and procedural memory (learned strategies).

The transient information an AI agent holds during a single task execution, analogous to a human's short-term memory. Working memory typically consists of the current context window contents, including the active conversation, retrieved documents, and intermediate reasoning results.

Persistent storage mechanisms that allow an AI agent to retain and recall information across separate sessions and conversations. Long-term memory is typically implemented via vector databases, knowledge graphs, or structured databases that the agent queries to inform current decisions.

A model compression technique that reduces the numerical precision of model weights from higher bit-widths (32-bit or 16-bit floating point) to lower ones (8-bit, 4-bit, or even 2-bit integers). Quantization dramatically reduces memory usage and inference cost, often with minimal quality loss, enabling large models to run on consumer hardware.

The process of transferring knowledge from a large, capable 'teacher' model into a smaller, faster 'student' model. Distillation trains the student to replicate the teacher's output distributions, reasoning patterns, and capabilities, producing a compact model that achieves most of the teacher's quality at a fraction of the inference cost.

The use of AI models to create artificial training data that mimics real-world data distributions. Synthetic data generation addresses data scarcity, privacy concerns, and class imbalance by producing large volumes of realistic, labelled examples for fine-tuning, evaluation, and testing.

The broad discipline of ensuring AI models behave in ways that are helpful, honest, and harmless — aligned with human values and intentions. Safety alignment encompasses training techniques (RLHF, Constitutional AI), evaluation (red teaming), and runtime safeguards (guardrails, content filtering).

A decoding strategy (also known as Top-P sampling) that dynamically selects the smallest set of tokens whose cumulative probability exceeds a threshold P. Unlike fixed Top-K sampling, nucleus sampling adapts the candidate pool size based on the model's confidence — using fewer candidates when the model is certain and more when it is uncertain.

The collection of techniques and architectural patterns designed to prevent adversarial users from hijacking an LLM's behaviour through crafted inputs. Defences include input sanitisation, instruction hierarchy enforcement, sandwich prompting, output monitoring, and privilege separation between system and user contexts.

Techniques that constrain a model's token-by-token generation process to produce output conforming to a formal grammar or schema. Unlike post-hoc parsing, structured generation enforces validity at decode time using context-free grammars, finite state machines, or schema-guided decoding.

A training methodology that fine-tunes a language model to reason over retrieved documents, including both relevant and irrelevant (distractor) passages. RAFT teaches the model to identify useful information within noisy retrieval results and cite sources accurately, improving RAG pipeline quality.

A prompting technique where the model generates an initial response, then creates and answers verification questions about its own claims, and finally revises the response based on the verification results. CoVe systematically reduces hallucinations by forcing the model to fact-check itself.

A technique that combines the weights of two or more fine-tuned models into a single model without additional training. Model merging methods (such as TIES, SLERP, and DARE) can create models that combine capabilities from multiple specialised fine-tunes, such as coding ability and medical knowledge.

An inference acceleration technique where a small, fast 'draft' model generates candidate token sequences that are then verified in parallel by the larger target model. Speculative decoding can speed up inference by 2-3× without changing output quality, because the large model verifies rather than generates.

A generation technique that restricts the model's output at each token position to only valid options according to predefined rules, grammars, or schemas. Constrained decoding guarantees outputs conform to required formats (valid JSON, SQL, code) without relying on retry logic or post-processing.

An advanced RAG architecture where an AI agent autonomously decides when, what, and how to retrieve information — dynamically formulating queries, selecting data sources, evaluating retrieval quality, and iterating until sufficient context is gathered. Unlike basic RAG, the retrieval process is driven by the agent's reasoning rather than a fixed pipeline.

A security vulnerability where an adversary manipulates a model into revealing its hidden system prompt, confidential instructions, or internal configuration. Prompt leaking compromises intellectual property and exposes security-critical rules that attackers can then use to craft targeted bypasses.

An API parameter that allows developers to directly increase or decrease the probability of specific tokens being generated. Logit bias modifies the raw output logits before sampling, providing fine-grained control over vocabulary usage without changing the model's weights.

A document processing technique that splits text into segments based on semantic meaning and topic boundaries rather than fixed character or token counts. Semantic chunking produces more coherent retrieval units for RAG pipelines by keeping related information together.

A security and design principle where different instruction sources (system prompt, user prompt, tool outputs, retrieved documents) are given explicit priority levels. The model is trained or prompted to always follow higher-priority instructions when conflicts arise, preventing user inputs from overriding system-level safety rules.