Guide to LLM resources
Understand each resource type: what it is for, when to use it and how.
Glossary: AI concepts explained simply →
Publish without the form: API, command line and GitHub Actions →
Use the catalogue in your tools
Your AI assistant (ChatGPT, Claude, Cursor, VS Code) searches and loads resources on its own.
pip install contexte: a dataset, a prompt or a context in one line.
All public routes, in JSON, no account: interactive documentation.
Personal token, ctx-publish, curl and GitHub Actions.
Production — run the model in production, without changing it
A context is a set of instructions and examples given to the AI model before the user's question.
A prompt is a reusable text template with variables in double braces, for example {{product}}.
An agent is an AI assistant that chains steps to complete a task: it follows instructions, uses MCP servers and contexts, and respects guardrails..
A RAG (retrieval-augmented generation) pipeline lets a model answer from your documents: they are chunked and embedded, then for each question the useful excerpts are retrieved and given to the model..
A workflow is a chain of steps where AI takes part: fetch data, have a model analyse it, then send the result elsewhere.
An output schema is a JSON Schema describing exactly the shape of the expected answer: which fields, which types, which are required.
Training — change the model by training it
A dataset is a set of examples in JSONL format (one JSON line per example) used to train or evaluate a model..
LoRA adapts an existing model to a task without fully retraining it: only small adapters are trained.
An evaluation is a set of test cases (input and expected answer) with a scoring method.
A corpus is a set of documents already split into chunks, each with its text and source.
A reward function scores a model's answer during reinforcement learning: the model learns to produce the answers that get the best score.
A synthetic data recipe explains how to generate a dataset with a model: the prompt that produces the examples, the model used, the target format, the volume and the filters that discard bad examples..
An annotation guide gives the instructions to label data consistently: the list of labels, the definition of each one and annotated examples.
Extensions — plug tools into the model
An MCP (Model Context Protocol) server gives an AI assistant new abilities: read files, query a database, call a service.
An A2A agent is an AI agent reachable through the A2A (Agent2Agent) protocol, an open standard run by the Linux Foundation.
A skill is an instruction pack (SKILL.md file) that an assistant like Claude loads only when the task calls for it.
A harness is the environment that runs an agent: its instructions file (e.g.
Memory lets an AI assistant remember information from one conversation to the next: preferences, decisions, project context.
Infrastructure — run a model on your own machine
An inference config is the set of settings needed to run a model on your own machine: the runtime (Ollama, vLLM, llama.cpp), quantization, context size and launch command..
A stack is a set of AI services that work together and start with one command: for example a model, a vector database and a chat interface, described in a docker-compose file..
Compliance — prove the AI is safe and lawful
A guardrail is a set of rules that controls what goes into or out of an AI assistant: block personal data, detect prompt injection, refuse off-topic requests..
Red teaming means deliberately attacking an AI assistant to find its weaknesses before users do: rule bypass, data leaks, prompt injection, promises it should not make..
An AI Act file is a documentation template to fill in to comply with the EU AI Act: transparency notice, training data description, risk assessment or system register..
An AI use policy sets an organisation's internal rules: which tools are allowed, which data must never be sent to an AI, when human review is mandatory.