ragleap-graph v0.9.0

[0.9.0]

Added

• Ontology cross-validation (#152). ExtractionConfig.relation_ontology (default None) lets callers define which relation_type values are semantically valid between which entity_type pairs, e.g. {"FOUNDED_BY": (["ORGANIZATION"], ["PERSON"])}. Requires entity_types= to also be set (raises ValueError at config time otherwise, matching the fail-fast style already used elsewhere in ExtractionConfig.__post_init__) - an ontology is meaningless without typed entities to check it against.
• LLMRelationExtractor.extract() gained an optional entity_types parameter (a {lowercased_name: type} map) threaded through from the entity_type_map already built during upsert_document()'s per-chunk loop - the type information existed in scope already, it just never reached relation extraction or validation before this.
• A relation whose relation_type is a key in relation_ontology, but whose subject/object entity types are not in that relation_type's allowed lists, is dropped (with a WARNING logged, not silently) rather than written to Neo4j. relation_type values NOT present in relation_ontology remain fully unconstrained - an ontology only needs to cover the relation_types worth constraining, not every possible one, mirroring entity_types='s own "unset means anything goes" precedent one level up.
• Applies uniformly to both the existing per-chunk extraction pass and the v0.8.0 cross-chunk pass, since both call the same LLMRelationExtractor.extract().

Verified

• 4 new offline tests: config-time ValueError when relation_ontology is set without entity_types, a relation matching allowed types is kept, a relation violating allowed types is dropped, an unlisted relation_type passes through unconstrained.
• 1 new live end-to-end test, gated on live Neo4j + GEMINI_API_KEY (same precedent as the #154 cross-chunk test: ontology validation depends on the underlying relation_type/entity_type extraction being accurate, which is unreliable on small local models). Confirms real Gemini-extracted output: an ontology-valid relation is kept, and a real ontology-violating relation the model actually proposed is genuinely dropped - not just asserted against mocked output.
• Full offline suite: 91 passed, 16 skipped without credentials (previous baseline 87/16 - 4 new tests added, zero regressions).