Glossary

The vocabulary used across these docs, in one place. Terms link to where they’re explained in depth.

chunk

A passage-sized piece of a source document — the atom of the corpus. See Corpus input.

facet

One of six independent slices the graph is projected into: foods, health, nutrients, dietary_patterns, allergies, sustainability. A chunk’s entity links route to the relevant facet(s). In the current corpus only foods is meaningfully populated (FoodOn is a food ontology); the others are scaffolded. See Layer A.

shelf

A node in the Layer A backbone — one FoodOn class kept because the corpus has evidence for it. Shelves form the coarse, browsable menu. See Layer A.

backbone

Used in two senses. (1) The ontology backbone: FoodOn itself, which the graph is projected onto. (2) The backbone projection (a.k.a. 1a+): the production Layer A method that picks a facet root’s supported children and expands down the real FoodOn tiers. Context disambiguates; this glossary entry exists because they collide.

1a+

The name of the production Layer A construction method (projection="backbone"). It comes from the method bake-off (method “1a”, plus refinements). Synonymous with backbone projection.

support (direct vs lifted)

The chunk evidence behind a shelf. Direct support = chunks that mention this exact FoodOn class. Lifted support = evidence inherited from descendants by walking up the is-a tree (a chunk mentioning cow milk lifts to mammalian milk product, dairy food product, …). A shelf with high lifted but ~0 direct support is an organizational umbrella; high direct support marks a genuine topic.

lifted attachment

Distinct from lifted support. A chunk attaches to every shelf on its lift path, so one chunk can sit on multiple shelves. (Lifted support is about a shelf’s evidence count; lifted attachment is about a chunk’s multi-shelf membership.) This multi-attachment is why Layer B themes are tied to an origin shelf.

umbrella class / umbrella rule

An organizational FoodOn class (food product, vertebrate food product) that absorbs generic mentions — high lifted, low direct support. The umbrella rule in the prune cascade drops such inflated classes. See Layer A.

filing tier

An intermediate is-a node that exists only to organize, with a single child and no navigational value. The backbone projection collapses these.

theme

A Layer B node — a fine-grained topic community discovered inside a shelf. See Layer B.

ThemeCandidate

A community emitted by one Layer B pass before merging. The merge turns candidates into final Themes. Each candidate carries its origin_shelf_id.

discovery_pass

How a theme was found: relatedness (entity pass only), global_similarity (embedding pass only), or merged (both agreed). Naming landmine: the value global_similarity is a historical name — in the production per-shelf mode it means per-shelf embedding similarity, not a global pass. See Layer B.

origin shelf

The shelf whose chunks a per-shelf theme was built from. A theme attaches to its origin shelf, not the union of its members’ shelves — avoiding cross-shelf smearing. See Layer B.

card

A Layer C node — a short, fully-cited LLM write-up for a shelf or theme. See Layer C.

NER

Named-entity recognition — finding food/health mention spans in text. FoodScholar uses GLiNER. See Annotation.

NEL

Named-entity linking — resolving a mention to a FoodOn id. FoodScholar uses a single-tier dense (HNSW) linker. Pre-computed NEL can be supplied as CSVs.

c-TF-IDF

Class-based TF-IDF: TF-IDF where each theme (not each document) is one “document”. Used to pick a theme’s discriminative keyword label.

RRF

Reciprocal-rank fusion — the method that blends BM25 and kNN result lists into one hybrid ranking at retrieval time.

denormalization

Copying a chunk’s shelf_ids / theme_ids onto its Elasticsearch document so retrieval can filter without round-tripping to Neo4j. Kept consistent by audit invariants. See Architecture.