Provenance
Lineage · Source trace
The recorded chain from an output back to its origin: which file, which page or time range, which extractor and which exact span produced it.
In DocSlurpGlossary entry
On this page
Provenance is the record of where a piece of output came from. For a document system it answers: which file, which version, which page or slide or timestamp, which region of the page, and which extractor turned pixels or bytes into this text.
Take the sentence “FY2027 revenue is projected at $48.2M.” Weak provenance says growth-plan-2027.pdf. Useful provenance says growth-plan-2027.pdf, page 12, characters 210 to 268 of the extracted text, a box covering the top third of the page, extracted from the PDF text layer by a named parser, from the version uploaded on a given date. With that, a reader can jump to the exact highlighted line, and an engineer can tell whether a wrong value came from the document, the extractor or the model.
Citations are provenance surfaced to the reader. Structured output needs it per field: a value of 30 is only checkable if you know which span said “30 minutes”.
The gotcha
Provenance can’t be reconstructed later. Every transformation, from layout analysis and chunking to embedding and summarizing, must carry its source pointers forward. A chunk that merged text from pages 11 and 12 without recording both, or a summary with no link to the passages it condensed, breaks the chain for good. Design the data model so every derived record holds a reference to what it was derived from.
In DocSlurp
In DocSlurp
Citations carry a typed locator (page with character offsets and normalized boxes, slide, sheet range, media time range, image region) and a sourceTrace naming the canonical document, parser artifact, block IDs and extractor (core/packages/shared/src/citation-locator.ts, core/worker/src/stages/structured-extract.ts).