LegalAI

Features

Ten capability areas, every one a live endpoint

Nothing here is a mockup. Each capability degrades honestly when a dependency isn't reachable.

Structured understanding

  • Clause & sentence segmentation into typed ClauseObjects
  • Deontic modality tagging — obligation / permission / prohibition / discretion
  • Defined-term extraction that doubles as party identification
  • Cross-reference resolution (“Section 4.2”, “Exhibit A”)
  • Money / jurisdiction entities via regex, merged with GLiNER zero-shot spans
  • Temporal expression normalisation — absolute dates resolved, durations left honest
  • Ambiguity & vagueness detection

Risk analysis

  • Keyword sweep across an 8-category taxonomy
  • Contextual AI risk pass on top of the rule floor
  • Per-category radar chart with clause-level drill-down
  • Cross-document conflict candidates from the knowledge graph

Grounded generation

  • Plain-English rewrite — whole document or a single clause
  • Contextualised explanation personalised to role / location / contract type
  • Free-form Q&A grounded in the actual contract text
  • Every answer entailment-checked; unsupported claims surfaced
  • Retrieved sources cited by a globally-numbered [N]; fabricated citations flagged

Agentic pipeline

  • A planner decides which agents run for this document
  • Risk & compliance, clause research (hybrid RAG), summary, verifier
  • Analysis modes: full / quick / risk-only / extract-only
  • Full step-by-step audit trace persisted per run
  • Verifier gate: citation validity + KG consistency + NLI faithfulness

Knowledge graph

  • Defined terms scoped per-document, linked across the portfolio when the context matches
  • “What else references this term” across every ingested document
  • Candidate cross-document obligation / prohibition conflicts
  • Bitemporal versioning — query the graph as of any point in time
  • Interactive node/edge explorer

Negotiation drafting

  • Clauses compared against your organisation's preferred language
  • Redline diff + rationale for every deviation — flagged on any change, not a similarity threshold
  • Always pending review; nothing is auto-applied
  • The Verifier folds suggestions into needs_human_review

Timeline & simulation

  • Document structure tree + descriptive timeline of dates and obligations
  • Discrete-event simulation over resolved absolute dates
  • Past / upcoming / future classification against a configurable warning window

Security & governance

  • Four-tier sensitivity classification enforced end-to-end
  • PII redaction gate before any external dispatch
  • SHA-256 egress audit trail
  • Per-key and per-user RBAC with a full audit log
  • Air-gapped build excludes the commercial provider package

Model router

  • Declarative task × sensitivity × capability routing policy
  • Class A / B / C hosting classes, self-hosted-first
  • Per-task Class C kill switches
  • Escalation ladder — bigger self-hosted model, never a bigger vendor

Evaluation

  • Graded harness over LegalBench / MNLI + a hand-curated gold set
  • Cutover gate: a task only becomes self-hosted-by-default when it meets the baseline
  • Self-hosted-vs-external delta report
  • Every run persisted for regression attribution