kayetanaGlossary
AI terms, explained simply.
40 terms around language models, privacy and operations.
- API
- An API is an interface that lets programs talk to each other. Instead of clicking in a browser, your app sends requests and gets responses back. kayetana offers a REST API under /v2.
- API key
- An API key is a secret key an application uses to identify itself to an API. Whoever has the key can make requests, so keep it safe. In kayetana, each key has its own monthly limit.
- Citation
- A citation shows which document a statement comes from. That lets you check any answer yourself. kayetana adds superscript numbers per sentence and a source list below.
- Classification
- In classification, AI assigns a text to one or more categories, such as sorting tickets by topic. Each label comes with a confidence score. In kayetana, you can classify using examples in the request or with a classifier you trained.
- Cloud AI
- With cloud AI, the models run in a provider's data center. Your input travels there over the internet. You don't need your own hardware, but your data leaves the building.
- Confidence
- Confidence shows how sure a model is about an assignment, usually as a value between 0 and 1. Low values are a hint to take a closer look. kayetana shows a confidence score per label when classifying.
- Context window
- The context window is the number of tokens a model can take into account at once: instructions, history, documents and answer combined. Anything that doesn't fit, the model doesn't see. kayetana shows the context length of every installed model.
- Cosine similarity
- Cosine similarity measures how closely two vectors point in the same direction. A value close to 1 means very similar meaning. That's how a search finds matching texts even when different words are used.
- Data processing agreement
- A data processing agreement applies when an outside provider processes personal data on your behalf. Under the GDPR, this usually requires a separate contract. When AI runs in-house, prompts don't go to an outside provider in the first place, though the legal assessment stays with you.
- Data sovereignty
- Data sovereignty means deciding for yourself where your data lives and who processes it. The less data goes to third parties, the more control you keep. With kayetana, prompts and documents never leave the building.
- Dataset
- A dataset is a collection of sample data, such as texts with labels or example conversations. It's used for embedding, training or testing. kayetana accepts datasets as CSV, JSONL, TXT, PDF or Word.
- Embedding
- An embedding is a piece of text turned into a list of numbers that captures its meaning. Texts with similar meaning get similar numbers. It's the foundation for semantic search, reranking and classification.
- Fine-tuning
- Fine-tuning trains an existing model further on your own examples so it better matches a style or task. The base knowledge stays, the behavior adapts. In kayetana, you fine-tune a chat model this way and get your own model named “ft-name”.
- Function calling
- With function calling, a model can call tools you define, such as a search or a database query. The model returns the name and arguments, and your app runs the call. kayetana supports tools in chat.
- Hallucination
- A hallucination is a statement that sounds convincing but is wrong or made up. Language models produce likely text, not verified facts. Source citations help you check answers.
- Inference
- Inference is using a finished model: you give it input and it computes an answer. Training happens beforehand, inference happens with every request. With kayetana, inference runs on your own Mac Studio.
- JSON mode
- In JSON mode, a model answers in valid JSON instead of free text. That way a program can process the answer directly. In the kayetana Playground, you turn on JSON mode per request.
- Language model (LLM)
- A large language model is an AI model trained on huge amounts of text that can understand and write text. It predicts, token by token, what most likely comes next. Chat, summaries and answers from documents all build on it.
- LoRA
- LoRA is a lightweight fine-tuning method. Instead of changing the whole model, it trains only small add-on weights. That needs less memory and compute, and kayetana uses LoRA to fine-tune qwen3:4b-instruct.
- Max tokens
- Max tokens caps how long a model's answer can get. Once the limit is reached, the answer stops. It keeps answers short and compute time in check.
- Multimodal
- A multimodal model handles more than text, for example images too. You can send a photo and ask questions about it. In kayetana, this works with models that support image understanding.
- On-premises
- On-premises means the software runs on your own hardware in your own building, not in the cloud. You keep control over the machine and the data. kayetana runs on a Mac Studio in your own office or network.
- Open-weight model
- With an open-weight model, the trained weights are freely available. You can download it and run it on your own hardware. Every model bundled with kayetana is openly licensed (Apache 2.0 or MIT).
- Parameters
- Parameters are the learned numbers inside a model. Their count, such as 4 or 20 billion, is a rough measure of size and capability. More parameters need more memory.
- Prompt
- The prompt is the input you give a model: a question, a task or a piece of text. The clearer the prompt, the better the answer usually fits. kayetana never stores prompts.
- Pseudonymization
- Pseudonymization replaces names and other direct identifiers with codes. Without extra information, a person can no longer be identified directly. It lowers risk but doesn't replace a careful review.
- Quantization
- Quantization stores a model's numbers at lower precision. The model gets smaller and needs less memory, often with only a small loss in quality. It lets larger models fit on ordinary hardware.
- RAG (retrieval-augmented generation)
- With RAG, the model receives relevant passages from your documents and answers based on them. That way the answer rests on your own files rather than just its training data. kayetana shows which source each sentence comes from.
- Rate limit
- A rate limit caps how much can be requested through an interface. Once the limit is exceeded, the API rejects further requests. kayetana sets a monthly limit per key and returns status 429 beyond it.
- Reasoning
- With reasoning, a model works through intermediate steps before it answers. That helps with tricky tasks but takes longer. In kayetana, you turn reasoning on or off and set a budget.
- Rerank
- Reranking sorts a list of documents by how relevant they are to a question. The most useful ones end up on top. kayetana does this using embedding similarity with bge-m3.
- Retention period
- A retention period sets how long data is kept before it gets deleted. Shorter periods mean less stored data. In kayetana, you choose unlimited, 7, 30, 90 or 365 days for usage logs and datasets.
- Self-hosting
- Self-hosting means running a service yourself instead of using it from a provider. You decide where it runs and who has access. kayetana is a self-hosted AI platform.
- Streaming
- With streaming, the answer arrives piece by piece while the model is still writing. You see the beginning right away instead of waiting for the end. kayetana streams chat answers via server-sent events.
- System prompt
- The system prompt is an instruction placed before the actual conversation that sets the model's behavior, such as its role, tone or rules. It applies to every message that follows. In the kayetana Playground, you set it as the system instruction.
- Temperature
- Temperature controls how random a model's wording is. Low values give consistent, factual answers, high values more variety. You can set it per request in the Playground.
- Token
- A token is a piece of text a language model works with, often part of a word or a short word. Models read and write text as a sequence of tokens. In kayetana, you can split text into tokens and back again on the Tokens page.
- Transcription
- Transcription turns spoken language into written text. Timestamps show when something was said. kayetana transcribes recordings such as MP3, M4A or WAV up to 100 MB.
- Vector
- A vector is an ordered list of numbers. In AI, a vector usually stands for the meaning of a text, that is, its embedding. kayetana returns vectors in float, int8, uint8, binary and ubinary formats.
- Whisper
- Whisper is an open speech recognition model for transcription in many languages. kayetana uses Whisper large-v3-turbo.