Skip to content

Extraction Profiles Reference ​

Detailed ontology reference for all five extraction profiles.

Profile comparison ​

Profileprofile= valueModelChunkingConfidenceEntity typesRelationship types
AEC / BIMaecFLAGSHIPtable_aware≥ 0.7566
Legal / ContractcontractFLAGSHIPclause≥ 0.7565
Financial / ERPfinancialDEFAULTtable_aware≥ 0.6865
ScientificscientificDEFAULTsection≥ 0.7076
GeneralgeneralDEFAULTparagraph≥ 0.65126

AEC / BIM (aec) ​

Entity types ​

IfcElement · IfcSpace · IfcMaterial · IfcSystem · RegulatoryCode · IfcClassification

Relationship types ​

MADE_OF · LOCATED_IN · CONTAINS · CONNECTS_TO · CLASSIFIED_AS · COMPLIES_WITH

Chunking: table_aware ​

Splits on paragraph boundaries but keeps tables intact. Table content is injected as dedicated LLM windows.


Entity types ​

Party · Clause · Obligation · Right · MonetaryValue · Date

Relationship types ​

SIGNED_BY · OBLIGATES · GRANTS_RIGHT_TO · GOVERNS · REFERENCES

Chunking: clause ​

Respects legal section / clause boundaries. Each numbered section is kept together.


Financial / ERP (financial) ​

Entity types ​

Company · Transaction · GLAccount · Amount · InvoiceLineItem · Date

Relationship types ​

ISSUED_BY · RECEIVED_BY · POSTS_TO · REFERENCES · PAID_BY

Chunking: table_aware ​

Same as AEC — paragraph-split with tables injected as dedicated windows.


Scientific (scientific) ​

Entity types ​

Author · Institution · Dataset · MLModel · Finding · Method · Metric

Relationship types ​

AUTHORED_BY · AFFILIATED_WITH · TRAINED_ON · EVALUATED_ON · OUTPERFORMS · REFERENCES

Chunking: section ​

Splits on paper sections (abstract, introduction, methods, results, discussion).


General (general) ​

Entity types ​

Person · Organisation · Location · Product · Event · Policy · Concept · Date · Value · Reference · Document

Relationship types ​

HAS_DATE · HAS_VALUE · IDENTIFIED_BY · REFERENCES · RELATES_TO · PART_OF

Chunking: paragraph ​

Standard paragraph-boundary splitting. Lowest confidence threshold (≥ 0.65) for maximum recall.


Auto-classification ​

When no profile= parameter is specified:

  1. Keyword heuristic — first 3,000 chars scanned for 13–14 domain keywords per profile
  2. Threshold — ≥ 3 keyword hits → profile selected (no LLM call)
  3. LLM fallback — if inconclusive, a one-shot classification call determines the domain

Keyword examples ​

ProfileSample keywords
contractcontract, agreement, clause, obligation, party, indemnity, warranty
financialinvoice, payment, ledger, debit, credit, balance, remittance
scientificabstract, methodology, hypothesis, dataset, findings, p-value
aecifc, storey, beam, column, concrete, HVAC, building

Confidence calibration ​

All profiles apply confidence calibration (temperature scaling T=2.0) after extraction. The per-profile threshold operates on calibrated scores, not raw LLM output.

Purple8 DocIntel v0.12.5 · Universal Document Intelligence · NVIDIA Inception Member · Red Hat Technology Partner.