What you can do today

  • Create private collections and typed entities, then find them by exact identifiers or literal text.
  • Record typed values, dated relationships and explicit proposed, approved, contested or revoked states.
  • Attach verified document words and page references; changed or deleted sources are excluded from returned knowledge.
  • Save named-field schemas as versioned definitions and reuse them on processed documents.
  • Turn an extracted field into a source-linked proposal for review; manual values remain labeled user assertions.
  • Owners grant existing verified accounts reviewer or viewer access to knowledge, feedback and read-only source evidence. Reviewers can propose corrections; owners control approval and access.
  • Correct facts and relationships with before-and-after values, reasons and revision history. Accepted replacements supersede the previous assertion.
  • Correct & learn records approved, currently effective organization context and invalidates qualification when relevant source evidence or approved context changes.

Correct a supplier record without losing its history

Create a supplier and attach the document words supporting its payment terms. A reviewer can propose corrected terms or an account relationship with a reason; an owner reviews the replacement. Keep the previous assertion and revision history available, and requalify learning rules if approved context changes.

Supported scope and limits

  • Search uses exact or literal matching. Semantic embedding collections and GraphRAG are not available.
  • The relationship view is a typed relational knowledge graph. A workflow DAG describes execution, while a lineage graph records source dependencies.
  • Selecting evidence or approving a record does not prove the underlying claim. Conflicts and stale facts still require judgment.
  • Shared access is limited to knowledge, feedback and read-only evidence; it does not grant shared invoice processing, billing or other owner-only operations.
  • Feedback learning supports context containing up to 100 approved knowledge records. Larger graph contexts are not yet supported for learning; knowledge editing remains available.
  • Automatic entity merging and organization-wide synchronization remain planned.

Planned extensions

  • Versioned embedding collections and hybrid retrieval over authorized evidence.
  • Explicit entity merge/unmerge decisions and wider source synchronization.
  • Changed-source impact across downstream workflows and derived summaries.
  • Graph-assisted investigation evaluated against simpler exact and lexical retrieval.

How to check the result

  • Inspect the original source before approving a fact.
  • Use explicit identifiers for entity identity; similar names are not a confirmed match.
  • Check effective dates, conflicts and current source availability before reuse.