KI-Konzepte

Standard pricing unit for AI APIs — cost in US dollars per one million processed tokens. Listed separately for input, output and sometimes cache.

An adapter is a small, freshly added network module injected into a frozen language model — only these modules are trained, while the base model itself stays unchanged.

Modular, reusable instruction packages (e.g. Claude Agent Skills) that an AI agent loads from disk on demand to perform specialized tasks — built around a SKILL.md holding metadata and instructions, plus optional scripts and reference files.

Agentic RAG is a RAG architecture in which an agent drives retrieval — it decomposes the question, searches iteratively, checks the hits, and decides itself whether a second or third pass is needed.

An AI agent combines a language model with tools and works toward a goal across multiple steps — typically in a loop of observing, planning and acting.

Computational method in transformer models that weights which parts of the input text are most relevant for predicting the next token — instead of treating every part as equally important.

Audio normalization is the adaptive linear scaling of an audio signal to a target level — the tool first measures a reference value (peak, RMS or LUFS) and from that calculates the required gain.

AutoGPT is an open-source experiment released in 2023 that runs a language model such as GPT-4 in an autonomous loop — it breaks a goal into subtasks itself, calls tools and works through them with minimal human intervention.

Automatic Speech Recognition (ASR), also called Speech-to-Text (STT), is the automatic conversion of spoken language into written text by an acoustic model.

Asynchronous API mode that collects many requests and processes them at a significant discount — results are typically delivered within 24 hours.

Batch inference is the bundled processing of many prompts in one pass — locally on a GPU or as an asynchronous API job — trading real-time latency for throughput and cost.

An AI benchmark is a standardized task suite that makes language models comparable — same prompts, same scoring, transparent results for reasoning, knowledge, or coding.

Browser use refers to AI agents that drive a real web browser — navigating, clicking, filling forms, reading content — to complete web tasks autonomously. The browser serves as both the tool and the perception surface.

Catastrophic forgetting describes the effect that a neural network loses previously acquired knowledge — wholly or partially — when it is trained further on new data; a key problem when fine-tuning language models.

Chain-of-Thought (CoT) is a prompting technique that asks the model to spell out its reasoning in intermediate steps — boosting accuracy on multi-step tasks.

Chunking is the splitting of longer texts into smaller, self-contained pieces — the first step of any RAG pipeline because embeddings and retrieval work at the chunk level.

US law from 2018 (Clarifying Lawful Overseas Use of Data Act). US authorities can compel US providers to hand over data even when it is stored outside the US — for example in the EU. It conflicts with the GDPR.

Computer Use is a capability of modern AI models that lets them operate a computer like a human — see the screen, move the mouse, use the keyboard — to perform tasks across arbitrary applications without an API.

Constitutional AI is a training method developed by Anthropic in which an AI model critiques and revises its own behaviour against an explicit list of principles (a "constitution") — rather than relying solely on human ratings.

Context engineering is the discipline of deliberately curating and maintaining everything inside the LLM context window — system prompt, tool definitions, RAG hits, memory, conversation history.

Context Precision is a RAG evaluation metric (0–1) measuring how well the retriever ranks relevant contexts above irrelevant ones — the focus is ranking quality, not raw hit count.

Context Recall is a RAG evaluation metric (0–1) measuring how completely the retriever supplied the facts needed for the gold answer — every missing fact lowers the score.

The context window is the maximum number of tokens a language model can process at once — input and output combined.

Contextual Retrieval is an Anthropic RAG technique that prepends an LLM-generated context sentence to every chunk before indexing — sharply raising retrieval quality.

Continued pretraining is the further training of an already pretrained language model on large amounts of domain text — before any classic fine-tuning begins.

Cosine similarity measures how similar two vectors are by the angle between them — the default metric in RAG retrieval for comparing query embeddings against the embeddings stored in a vector index.

Processing on behalf occurs when a service provider processes personal data for a controller, bound by instructions. The Data Processing Agreement (DPA) under Art. 28 GDPR governs the mandatory terms of this arrangement.

Data residency refers to the physical storage and processing location of data — the geographic region or country where the servers sit. It says nothing about which legal jurisdiction may access that data.

Three related concepts for digital audio levels: dBFS is the scale with 0 dBFS as the hard ceiling, headroom is the safety margin below the maximum, clipping is the hard cut-off when a signal exceeds it.

DPO is a fine-tuning method that aligns language models directly to human preference pairs — without a separate reward model and without reinforcement learning.

An embedding is a numeric representation (vector) of text, an image, or other data in which semantically similar content sits close together in space.

Ensemble means combining several deliberately varied LLM runs or models whose findings complement each other. Multi-model orchestration drives these runs via orchestrators with sub-agents, so the union of results is larger than any single run.

Regulation (EU) 2024/1689 is the EU's first comprehensive law on artificial intelligence. It regulates AI systems by risk in four tiers, from prohibited practices to minimal risk, and sets obligations for providers and deployers.

Evals are systematic tests for LLM applications — fixed test cases, automatic scoring of the responses, result as a score. The foundation for objectively comparing prompt or model changes.

Faithfulness is a RAG evaluation metric (0–1) measuring how many claims in a response can actually be derived from the retrieved contexts — a direct hallucination indicator.

Few-shot prompting is a technique that includes a handful of input/output examples in the prompt so the model picks up the desired format and style.

Continued training of a pretrained model on a smaller, specialized dataset to adapt style, format or domain knowledge in a targeted way.

Function calling is a language model's ability to produce a structured function invocation instead of a text reply — the technical foundation for tool use and AI agents.

GDPR-compliant LLM use describes the data-protection-compliant use of large language models such as ChatGPT, Claude or Gemini in a company — with a legal basis, data processing agreement, controlled input of personal data, EU hosting and training opt-out.

GGUF (GPT-Generated Unified Format) is the standard file format for quantised LLMs on the llama.cpp engine — a single `.gguf` file holds weights, tokenizer and metadata together.

GraphRAG is a RAG variant that uses a knowledge graph in addition to vector search — answers come from linked entities and relationship paths, not just similar text.

A reference taken to be correct, against which model outputs are measured. In AI code analysis it is a piece of code with deliberately injected, known defects, so you can measure how many real problems a model actually finds.

Guardrails are protective layers around a language model that inspect inputs and outputs to catch unwanted behaviour — off-topic answers, PII leaks, unsafe actions.

A hallucination is a language-model response that sounds plausible but is factually wrong or fabricated — typically caused by statistical guessing without a factual anchor.

Hybrid search combines lexical full-text search (e.g. BM25) with semantic vector search and unites the strengths of both — precise on keywords and robust on meaning.

HyDE (Hypothetical Document Embeddings) is a RAG technique where an LLM first drafts a fake answer to a query — and then uses that answer's embedding to find real documents.

Indirect prompt injection is an attack that hides malicious instructions inside external content — web pages, documents, emails — that an LLM then processes and unintentionally executes.

Inference is the act of running an already-trained model to turn an input into a response — production use, not training.

Tokens you send to an AI model in an API call — your prompt, the context, attached documents. Billed separately from output tokens and usually much cheaper.

Instruction tuning is a variant of supervised fine-tuning in which a language model is trained on a broad mix of instruction-response pairs — so it reliably follows instructions across arbitrary tasks.

A jailbreak deliberately bypasses a language model's safety and behaviour rules through cleverly crafted prompts, making the model produce content it would normally refuse.

A keyword glossary for Speech-to-Text is a short, curated list of project-specific terms, brand and product names handed to an STT model as a context hint before transcription — typically through Whisper's `initial_prompt`.

Knowledge distillation is a training technique in which a small student model learns from the behaviour of a large teacher model — aiming for comparable quality at significantly lower resource cost.

The KV cache (key-value cache) stores the already computed key and value vectors of a language model's attention layers during text generation. This avoids recomputing the entire context for each new token and significantly speeds up inference.

LLM stands for Large Language Model — a neural network trained on large volumes of text to interpret and generate natural language.

LLM-as-a-Judge is the practice of using a language model as an evaluator — it compares responses from other models or scores outputs against given criteria.

Low-Rank Adaptation — parameter-efficient fine-tuning that trains only small additional matrices instead of all model weights.

MCP — Model Context Protocol — is an open standard from Anthropic for connecting AI models to external data sources and tools through a unified interface.

Meta-prompting is the technique of having an LLM write or improve a prompt for a task rather than formulating it yourself — the prompt becomes the output.

A model card is a standardised datasheet for an AI model — documenting purpose, training data, performance, limitations and ethical considerations in one place.

ORPO is a training method that unifies supervised fine-tuning and preference optimisation in a single step — leaner than the classic SFT-plus-RLHF pipeline.

Tokens an AI model produces as its response. Billed separately and usually three to five times more expensive than input tokens because the model has to actively generate them.

Path dependency describes how, during LLM generation, each early token shapes everything that follows — the first verbalized finding steers the entire remaining analysis in one direction.

PEFT bundles methods that adapt a pretrained LLM by training only a small fraction of its weights — typically under 1%, instead of all billions of parameters.

Prefix tuning is a PEFT method that trains a sequence of learned "virtual tokens" prepended to the input rather than any model weights — the base model itself stays frozen.

Prompt caching is an API feature in which a provider stores recurring prompt prefixes — making subsequent requests cheaper and faster because the cached portion is not reprocessed.

Prompt engineering is the discipline of steering AI models through carefully crafted input prompts — closer to precise writing than to classic programming.

Prompt injection is an attack that hides instructions inside an LLM's input data to make the model ignore or subvert its original instructions.

Prompt leaking is an attack that tricks an LLM into revealing its hidden system prompt or other confidential context contents — a special case of prompt injection.

A prompt template is a reusable prompt skeleton with placeholders that get filled with concrete values at runtime — the basis for reproducible LLM calls in applications.

Extension of LoRA that quantizes the base model to 4 bits — allowing even very large LLMs to be fine-tuned on a single GPU.

Quantization reduces the numeric precision of model weights — e.g. from 16-bit float to 4-bit integer — making language models small and fast enough for consumer hardware, with manageable quality loss.

Query expansion enriches the original query in a RAG pipeline with synonyms, reformulations or hypothetical answer texts in order to lift retrieval recall.

RAG (Retrieval-Augmented Generation) connects a language model to an external knowledge source — relevant passages are retrieved and passed to the model alongside the question.

RAGAS is an open-source framework for automated evaluation of RAG and agent pipelines — with built-in metrics like faithfulness, context precision, and answer relevancy.

Provider-enforced cap on requests or tokens per time window — it protects infrastructure and ensures fair usage across customers.

ReAct is a prompting pattern that has an LLM alternate between reasoning (thoughts) and action (tool calls) — the foundation behind many agent implementations.

Reasoning Effort is a control parameter on modern reasoning models that sets how much internal step-by-step thinking (thinking tokens) a model spends before answering — higher levels raise quality but also latency and cost.

Recall (hit rate, sensitivity) measures the share of all actually existing relevant cases that a system finds. Formula: relevant cases found divided by all relevant cases that truly exist. Requires a known ground truth.

Reranking reorders an already-retrieved list of hits with a more accurate model — typically a cross-encoder that scores query and each candidate together, instead of just comparing vector distances.

Reinforcement Learning from Human Feedback — training procedure that aligns models with helpful and safe behavior using human preference comparisons.

Role prompting assigns the LLM a concrete role or persona ("you are an experienced tax lawyer …") to steer style, vocabulary and depth of answers.

S-RAG (Search-optimized RAG) is boostN.ai's take on Retrieval-Augmented Generation. It extends classic RAG with search-engine principles: relevance ranking, recency decay and typed links between chunks for deterministic follow-up retrieval.

Safetensors is a file format developed by Hugging Face for storing model weights — faster, safer and more language-agnostic than the older PyTorch `.pt`/`.bin` format.

Sampling is the weighted drawing of the next token from a language model's probability distribution — governed by temperature, top-p and top-k. It produces the variability between two runs.

Self-Consistency is a prompting technique where the same question is answered multiple times with chain-of-thought, and the most frequent answer is picked as the final result.

Self-Refine is a prompting technique in which a model critiques its own output and improves it across several iterations — without human feedback.

Semantic search retrieves content by meaning rather than exact word match — query and documents are compared as embeddings, so synonyms and paraphrases also hit.

SFT is the supervised post-training of a pretrained language model on a dataset of input-output pairs — the step that turns a base model into a usable assistant.

Spotlighting is a defense technique against prompt injection in which untrusted inputs are marked so the model treats them as data — not as instructions.

Stop sequences are strings that make an LLM end token generation as soon as they appear. They constrain the output deliberately — to prevent role switches, format markers or overly long answers.

Streaming means transmitting an LLM's response token by token in real time — the user sees the text word by word rather than only after the full generation completes.

Structured output is a language model's ability to deliver responses in a defined schema (typically JSON), reliable enough to be consumed directly by downstream code.

Synthetic data is artificially generated training or evaluation data — typically produced by a language model itself to extend datasets without needing real-world sources.

A system prompt is the instruction that shapes a language model's behavior across an entire conversation — in contrast to the user prompt, which varies per message.

Temperature is a sampling parameter that controls how deterministic or creative a language model's responses are — low values stay conservative, high values produce more variety.

The thinking budget (token budget) is the maximum number of internal reasoning tokens a reasoning model may spend on "thinking" per request — controlling the trade-off between answer depth and cost or latency.

Tiered model line-up from a provider — small fast variants (Mini/Flash/Haiku) at a fraction of the price of the big frontier models. Also: volume tiers with quantity discounts.

A token is the smallest unit a language model works with internally — usually a sub-word fragment, occasionally a single character.

A tokenizer is the program that splits text into tokens before a language model can process it — it determines how many tokens a piece of text costs.

An invocation of a tool by an AI model during a conversation — such as reading a file, running a bash command, fetching from the web or calling an MCP tool. The foundation of agentic workflows.

Top-p (nucleus sampling) and top-k are sampling strategies for LLMs that decide which token candidates to draw from at each step — together with temperature they control the creativity of the output.

TPM (tokens per minute) and RPM (requests per minute) are the two common units in which AI API providers express their rate limits — TPM caps the token volume, RPM caps the number of requests per minute.

Tree of Thoughts (ToT) is a prompting technique in which an LLM explores multiple solution paths as a branching tree, scores them and follows only the promising branches — a generalisation of Chain-of-Thought.

A vector database stores embeddings as high-dimensional vectors and searches them by semantic similarity — the core infrastructure for RAG, semantic search, and recommendation systems.

VRAM (Video RAM) is the dedicated memory on a GPU — the single most important hardware figure for running language models locally, because ideally the whole model and its context must fit inside it.

Whisper is OpenAI's open speech-to-text model from 2022 — a multilingual encoder-decoder transformer that turns audio into text and ships in several sizes under an MIT license.

The `initial_prompt` parameter lets you pass Whisper a keyword list or sample sentence before transcription. The model treats this text as context and recognises the included terms far more reliably.

XML tags are named brackets like <context>...</context> used to clearly separate sections of a prompt. They structure the input, split instruction from data and make the model's output more predictable.

Zero-shot describes solving a task with a language model without providing any examples — purely via the task description in the prompt.