bobinet

RDF knowledge graphs from text, with the source sentence behind every statement.

Give Bobinet a Markdown document and it writes a TriG graph. It uses Claude to extract the entities and relations, published ontologies supply the terms, and entities link to Wikidata and DBpedia.

It's initial use case was in support of the Thick Edges demo (forthcoming publication) vocabulary covers intellectual history: people, works, ideas, and who influenced whom.

Bobinet follows the four stages of Anthropic's knowledge-graph cookbook: extract, resolve, assemble and query.

The name comes from bobbinet, a machine-made net.

Requirements

• Python 3.10 or later: pip install -r requirements.txt
• the claude CLI on your PATH, logged in. Bobinet calls Claude through it.
• network access to DBpedia, Wikidata and the ontologies' servers

Run

python bobinet.py --src note.md --out out all                     # build the graph
python bobinet.py --out out ask "Who influenced whom, and how?"   # answer from the graph

all runs seven stages, and each also runs alone:

stage	what it does
chunk	split the document at ## headings
extract	entities and relations per section (Haiku)
resolve	merge the names that refer to one entity (Sonnet)
link	link entities to DBpedia, then Wikidata (Sonnet)
assemble	write <name>.trig and <name>.ttl
validate	check shapes, terms and quotes into validation-report.md
query	run queries/*.rq into query-results.md

An interrupted extract resumes where it stopped. To redo part of a run:

python bobinet.py --out out extract --chunks s05,s06        # redo two sections
python bobinet.py --out out extract --force                 # redo all sections
python bobinet.py --out out link --entities "Carl Menger"   # re-link one entity

Evidence

For every relation, Claude copies the sentence that supports it. Bobinet looks for that sentence in the section, ignoring Markdown markup, spacing and typographic punctuation. If the sentence is not there, Claude gets one more try. A claim still without its sentence is kept and marked kgx:evidenceVerified false.

Each relation sits in its own named graph, so the provenance graph can describe it. The graph is named after the document, the section and the relation's position in that section:

claim:hayek-lines-of-influence.s08.r1 {
    dbr:Milton_Friedman  dbo:influencedBy  dbr:Friedrich_Hayek .
}

# in the provenance graph
claim:hayek-lines-of-influence.s08.r1  a kgx:Claim ;
    kgx:confidence       0.90 ;
    kgx:evidenceVerified true ;
    prov:value           "What Friedman took from Hayek here was not a theorem. It was a demonstration of method" ;
    prov:wasQuotedFrom   <.../doc/hayek-lines-of-influence/section/s08> ;
    prov:wasGeneratedBy  <.../activity/run-20260825T095909Z> .

With a name for each statement, you can:

• show the evidence for any edge
• drop one statement and keep the rest
• merge graphs from many documents and still tell which document said what

For tools without named graphs, <name>.ttl holds every statement in one graph. It keeps the quotes but not which statement each belongs to.

Cost

Every run spends Claude tokens. Bobinet needs Claude to extract relations in a closed vocabulary, each with its quote. Linkers such as DBpedia Spotlight and TextRazor spend no tokens. They find and link names, and take their relations from a vendor's fixed list or from grammar.

A run makes one Claude call per section, a retry where a quote fails, and up to eight calls for merging and linking.

Entity linking

A candidate links only when its registry description agrees with the document's description of the entity. So a document about the jazz pianist John Hicks does not link to the economist of the same name.

Bobinet tries DBpedia first, then Wikidata. Each verdict is matched, rejected, or unsure. links.json records the verdict, the reason and the candidates weighed.

Vocabulary

Claude may use only the 6 entity types and 19 relations in profile.ttl, reused from schema.org, FOAF, SKOS, DCMI Terms, PROV-O, CiTO, FaBiO, DoCO and the DBpedia ontology. Bobinet drops any relation outside them. It mints only the three kgx: terms.

To extract for another domain, edit the predicates in prompts.py and their IRIs in bobinet.py, then update profile.ttl and shapes.ttl to match.

Options

flag	meaning
--src	the Markdown document; required by chunk
--out	output directory, default ./out
--name	basename of the .trig and .ttl, default the document's slug
--base	base IRI of the run's data, default https://personalknowledgegraphs.com/data/bobinet/
--iri-policy	wikidata (default), dbpedia or local: whose IRI names a linked entity
--gold, --gold-entity	a DBpedia SPARQL result with one person's influence lists, and that person's label; validate reports the overlap

Unlinked entities get IRIs under <base>id/ whatever the policy. The dbpedia policy also uses DBpedia-ontology properties and classes where an equivalent exists. To switch policy, re-run assemble, validate and query.

Zoom

zoom unfolds a claim into the finer statements behind it, each with its own quote and named graph. all does not run it.

Today zoom unfolds influence claims only. It finds the ideas one person took from another, rejected, reformulated, extended or disputed, and writes them with the fr: vocabulary at http://velitchkov.eu/def/fr#.

python bobinet.py --out out zoom --chunks s05,s08

Without --chunks, zoom takes every section that an influence claim quotes. It writes <name>-zoom.trig and <name>-zoom.ttl and leaves the other files alone. Nodica displays them with click and wheel zoom.

Roadmap

• Vocabulary profiles for any domain, with a general profile as the default
• Zoom for any relation a profile names
• The position of each quote in the source
• A quote for dates and other values
• Plain text and HTML input
• Entity linking scored against DBpedia Spotlight and TextRazor
• Candidates from a local linker, and a choice of model, to lower the cost

License

MIT. See LICENSE.