Skip to content

Upload time is used as the document date in extraction #714

Description

@WaylandYang

An undated upload stores doc_time = now() with doc_time_source = 'upload_time' (crates/utopia-store/src/documents.rs). That time is when the ledger learned of the document, the transaction axis. Only crates/utopia-store/src/temporal.rs filters on doc_time_source; extraction uses doc.doc_time as if it were the document's own date:

  • crates/utopia-server/src/extraction.rs passes it to the prompt, where crates/utopia-extract/src/lib.rs tells the model "Document date: {t}. Resolve relative time expressions … using it as the reference." A "today" or "last year" in an undated document resolves against the upload day.
  • validity_of sets attested_at: doc_time, so facts from an undated document are attested at the upload moment, and open-started facts hold from it.
  • name facts are recorded with doc.doc_time as their time.

Seen on 2026-09-15: every document in the recall-bench KB has doc_time_source = 'upload_time' and doc_time equal to the upload day; a prototype that passed it as "dated" produced "the 2024 statistical bulletin — dated → 2026-09-14".

Expected: a doc_time whose source is upload_time is treated as no document date in extraction (prompt says the date is unknown, no relative resolution against it, attested_at not taken from it). The document's own date should come from the source or from its content.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions