Jargon lookup
The nuggets try to gloss terms as they go. This page is the quick lookup: one plain sentence each, with a link to the nugget where the idea is developed.
RAG (retrieval-augmented generation): the common setup: upload documents, retrieve matching chunks at query time, generate an answer. The model rediscovers knowledge on every query; nothing accumulates. → Karpathy’s llm-wiki
llm-wiki: instead of retrieving from raw files, an LLM builds and maintains a wiki as it works, so knowledge compounds. → Karpathy’s llm-wiki
Front matter / typed edges: the small YAML block at the top of each page (type:, up:, extends:, supports:, criticizes:, related:) that records how pages relate. → Wiki front matter
RDF triple: a fact stored as subject-predicate-object, like this page criticizes that page, in a form a database can query. → The wiki is a knowledge graph
Reify: to turn something implicit (a typed edge in prose) into an explicit data record (an RDF triple). → Wiki front matter
SPARQL: the query language for RDF graphs; the graph equivalent of SQL. → The wiki is a knowledge graph
SHACL: a rule language that checks a graph’s structure is valid (for example, every supports: must cite a source). → The wiki is a knowledge graph
Ontology: an explicit statement of what the entities are, how they relate, and what the fields mean. Kept deliberately small here. → Ontologies, kept small
Random-Walk-With-Restart (RWR): a graph-ranking method that spreads importance across a graph from a set of starting nodes; the idea behind PageRank. → The wiki is a knowledge graph
Knowledge bundle: a wiki authored so that other people’s LLMs can consume it; the shareable unit of the ecosystem. → Knowledge bundles
OKF (Open Knowledge Format): Google Cloud’s open specification (2026) that formalizes the llm-wiki pattern as a portable bundle of markdown with YAML front matter, so wikis by different producers work with different agents. → Knowledge bundles
Agent card: a small public profile a group’s agent publishes: what it knows about and how to reach it. → Agent federation
ask primitive: the federation’s query channel: find which group has relevant knowledge, then query it (with permission). → Agent federation
MCP (Model Context Protocol): an open standard that lets any AI host (Claude Desktop, Cursor) use a shared tool or read shared data. → Agent federation
Connector: an agent that wraps a data source, declares a capability, does the interpretive extraction, and files results with provenance. → Connectors are agents
Marketplace: the distribution-and-discovery layer for installable tool plugins, the way federation is for bundles. → A marketplace for tools
Broker pattern: a non-LLM component holds the credential and returns only data, so the agent never sees the secret. → ControlMaster
ControlMaster: an SSH feature that authenticates a connection once and reuses it; a working instance of the broker pattern. → ControlMaster
Data fabric: an AI-driven integration layer that connects scattered data sources into one queryable, data-agnostic environment. → Data fabrics
FAIR: data that is Findable, Accessible, Interoperable, and Reusable; the data-management standard funders increasingly require. → Data fabrics
DID (decentralized identifier): a persistent identifier its subject controls cryptographically, resolvable without any central registry. → Decentralized identifiers
did:web: the simplest DID method: the identity document is just a file on a normal web domain, no blockchain. → Decentralized identifiers
Verifiable credential: a cryptographically signed claim, like a digital ID card, checked against a DID. → Decentralized identifiers