Skip to content

Concepts Glossary

Jump to bottom
Chris Sweet edited this page Jul 14, 2026 · 1 revision

type: reference up: "What-is-the-llm-wiki-Template" related:

Concepts Glossary

Quick-reference definitions for concepts a user encounters in the template's documentation and in the community discussion around it. Retrieval-theory and knowledge-management terms are grounded in the vision wiki; format and tooling terms in scripts/kg/README.md and the template's own docs. Terms marked "vision-wiki framing" are part of the reasoning trail behind the architecture, not shipped template mechanics.

Retrieval and graph concepts

Markov chain (in retrieval) — A model of a walk over a graph where the next node depends only on the current node and the edge weights. In the vision-wiki framing, retrieval is a Markov walk over the wiki/corpus graph that returns a ranked neighborhood rather than an exact match, so it degrades gracefully as structure thins, unlike a crisp query that returns empty on sparse structure.

Random walk with restart (RWR) — A Markov walk that, at each step, either follows an edge or "restarts" back to the seed node(s) with some probability. This biases the stationary distribution toward the neighborhood of the seed, ranking nodes by proximity. It is the primitive the vision-wiki architecture uses at both scales (wiki pages and corpus chunks). Restart probability and edge weights are tuning parameters. (Tong, Faloutsos, Pan, 2006.)

Personalized PageRank — PageRank (Page, Brin, Motwani, Winograd, 1999) computes a node's global importance from the link structure; the personalized variant seeds the walk from specific nodes so the ranking is relative to a query rather than global. Mathematically equivalent to RWR; the template's lineage leans on it as the retrieval primitive.

Entity resolution — Collapsing different surface forms of the same entity ("Obama", "President Obama", "Barack Obama") to one canonical node. The template avoids heavy versions of it because it is unreliable on real corpora: under-merging scatters an entity across surface forms, over-merging collapses distinct entities ("Washington" as person, state, university). Lightweight extraction keeps surface forms distinct and lets the walk propagate mass across them via co-occurrence, so there is no resolution step to fail. See Template-Philosophy.

Structural graph vs content graph — A structural graph derives from layout and containment (chunks in documents, up: chains, wikilinks); its edges are unambiguous and cheap. A content graph derives from what the material means in a normalized ontology (resolved entities, formally typed cross-document assertions); it is more expressive but expensive and fragile. The template favors structural plus lightweight extracted content over heavy content-graph engineering. See Knowledge-Graph.

Vanilla RAG vs graph-informed retrieval — Vanilla RAG embeds chunks and retrieves by dense similarity to the query; excellent for single-hop lookups, but it rank-collapses on deep multi-hop questions because it cannot traverse the alternate views the embedding does not encode. Graph-informed retrieval walks an explicit node-edge structure, assembling the multi-hop context envelope. The template treats them as complementary: dense retrieval as the default where it fits, graph-informed retrieval for evidence assembly. See When-to-Use-When-Not.

Context envelope (vision-wiki framing) — The local neighborhood of concepts, layout, and cross-document evidence around a question. The claim: RAG is really about assembling this envelope, and the model cannot answer without it even when the correct chunk is in the top-K.

Retrieval-hop vs reasoning-hop (vision-wiki framing) — A reasoning-hop is a step in decomposing the question; a retrieval-hop is a step in assembling the evidence on the index side. Question decomposition addresses reasoning-hops but does not by itself assemble the retrieval-hops, which is why dense retrieval fails on deep multi-hop questions.

Topology question vs content question — A topology question asks what connects to what (multi-hop, hub detection, edge-typed traversal); answer it with the knowledge graph via SPARQL. A content question asks what a page actually says (definitions, claims, numbers); answer it by reading the file. The template's core retrieval guidance: use the graph to find where to look, then read to learn what it says.

Knowledge-management and cognition concepts

Knowledge, expertise, experience (vision-wiki framing) — The three layers a well-curated wiki can capture: knowledge (facts, definitions), expertise (the author's judgment), experience (concrete cases, what was tried, what failed). Separable but not hierarchical. A wiki with all three is closer to a graduate advisor than a textbook. See Template-Philosophy.

DIKW pyramid — Ackoff's (1989) Data → Information → Knowledge → Wisdom hierarchy. The K layer maps loosely onto the wiki's knowledge layer; the vision wiki notes DIKW does not map cleanly onto its three-layer framing.

Tacit knowledge — Polanyi's (1966) knowledge that is hard to articulate: the ground of expert judgment ("we know more than we can tell"). Corresponds most closely to the experience layer.

Procedural vs declarative knowledge — Anderson's (1976) distinction: declarative knowledge is knowing that (facts); procedural knowledge is knowing how (skills, judgment applied in practice). The wiki's knowledge-vs-expertise axis echoes this distinction.

Deliberate practice — Ericsson and Charness's (1994) account of how expertise is actually acquired: gradual, effortful, feedback-driven practice. Cited in the template's framing to explain why the Matrix-style instant skill download stays science fiction (skill is embodied and gradual), even as the bundle-plus-LLM substrate approximates rapid competence acquisition another way.

Extended-mind tradition — The lineage (Bush's Memex, 1945; Licklider, 1960; Engelbart, 1962; Clark and Chalmers' The Extended Mind, 1998) holding that cognition legitimately extends into external artifacts an agent reliably uses. The template draws on it because a knowledge bundle plus a local LLM is exactly such an artifact: rapid competence acquisition via the receiver's workflow, not their neurology.

Memex — Vannevar Bush's 1945 proposal for a personal, curated knowledge store with associative trails between documents. The closest ancestor of the llm-wiki pattern; the part Bush could not solve (who maintains the trails) is what the LLM now handles.

Rapid competence acquisition — The practical goal the Matrix helicopter analogy names: a bundle consumer gains working competence in a domain fast, because the bundle ships judgment and experience alongside facts and the LLM walks it alongside the consumer's own notes. See Template-Philosophy.

Format and tooling concepts

Knowledge bundle — A portable, git-distributable, LLM-authored wiki on a topic, consumed through an LLM ask session. See Knowledge-Bundles-Overview.

OKF (Open Knowledge Format) — Google's minimal, permissive v0.1 standard for a directory of Markdown-plus-frontmatter cross-linked by Markdown links. Template import/export compatibility is planned, not yet shipped (see Knowledge-Bundles-Overview); the template's position is "OKF (the wire format) + the discipline layer (the contribution)".

RDF Turtle (.ttl) — A compact, human-readable text serialization of an RDF graph as subject-predicate-object triples. The template emits the graph as Turtle; rdflib reads/writes it natively and it loads into any triplestore. See Knowledge-Graph.

RDF-star — An extension of RDF that lets a triple be the subject or object of another triple, i.e. statements about statements. The template uses it to attach weights to mentions edges (graph-weights.ttl).

JSON-LD — A JSON serialization of an RDF graph. The KG extractor (wiki-to-jsonld.py) emits JSON-LD as its native output (graph.jsonld), which is then translated to Turtle.

SHACL (Shapes Constraint Language) — A constraint language for RDF: shapes describe what a conformant graph should look like (node types, required predicates, edge domain/range). pyshacl validates the extracted graph against fetched shapes.ttl and writes validation-report.ttl; the build reports violations but does not abort. See Troubleshooting.

SPARQL — The query language for RDF graphs. The template ships a curated set of canned queries under scripts/kg/sparql/ so you need not write SPARQL for common topology questions; you write your own only for uncovered topology patterns.

rdflib / pyshacl — The Python libraries the KG pipeline runs on, in-process (no Java, no server). rdflib parses/serializes RDF and runs the materialization CONSTRUCT queries; pyshacl runs SHACL validation. Note the pipeline does not use arq (Jena's SPARQL CLI); arq only appears on the optional Fuseki/Jena path.

Apache Jena Fuseki — An optional SPARQL server that can host the produced graph-full.ttl for multi-client query, SPARQL UPDATE, or federation. Opt-in; the default is in-process rdflib. See Knowledge-Graph.

Template and process concepts

Ingest / Query / Lint — The three operations the agent performs against the wiki: write new knowledge, read to answer, maintain health. See Everyday-Operations and Template-Philosophy.

Typed edge — A frontmatter relationship (extends:, supports:, criticizes:, source:, up:, related:, and more) that carries an interface contract: traversing it licenses a specific agent operation (inherit context, aggregate evidence, detect contradiction, ground a claim). Inverses are materialized by the KG, never authored. See Wiki-Structure-and-Conventions.

mentions edge — The generic body cross-reference ([Display](Page-Name)) the KG extractor emits automatically for any plain body link; weighted via RDF-star. Distinct from the semantically typed frontmatter edges.

Hub / orphan — A hub is a page with many inbound links (materialized as a flag by the KG); an orphan has no inbound links and is invisible to navigation and retrieval. Orphan detection is a standard lint check. See Wiki-Structure-and-Conventions.

Verification Gate — The pre-commit checklist every wiki write runs before staging (wiki/agents/verification-gate.md): corpus-tagged numbers, resolved links, paired back-references, valid frontmatter, updated index/log, honest reporting. See Everyday-Operations.

Discipline gates — The agent-agnostic enforcement patterns (wiki/agents/discipline-gates.md), including the "Universal Rationalizations (Always Wrong)" table, that convert honesty rules into pre-write checks. See Team-Workflow.

Wiki write protocol / wiki_push — The optimistic-push-with-retry wrapper for pushing a shared wiki sub-repo without collisions; union-merges index/log, defers content conflicts to the agent's next turn. Reference implementation shipped, not yet wired into the default write path. See Team-Workflow.

Agent overlay — The agent-specific surface layer (Claude Code slash commands + skills + hooks; Cursor rules; or --agent=none) on top of the agent-agnostic core. See Installation-and-Setup and Extending-the-Template.

SessionStart hook / PostToolUse hook — Claude Code hooks the overlay installs: SessionStart auto-clones and refreshes the wiki (ensure-wiki.py) and surfaces the index and recent log; the optional advisory PostToolUse hook nudges the agent through the verification gate after a wiki write. See Installation-and-Setup.

Template vs derived project — The template ships scaffolding (instantiate.sh, init-wiki.sh, overlays, policy files); a derived project is an instance with its own wiki/<repo>.wiki/, CLAUDE.md, and content. Template improvements flow to derived projects via update-from-template.sh and a file manifest. See Installation-and-Setup.

Feature (opt-in) — A self-contained, optional, removable extension under features/<name>/ with a feature.json manifest, enabled/disabled independently. First shipped feature: agent-comms. See Extending-the-Template.

See also

Clone this wiki locally