CORE GUIDE
Document extraction as evidence-preserving transformation
Document extraction should convert messy source material into structured evidence without erasing location, provenance or uncertainty.
Mental model
Extraction is a reversible transformation from source artifacts to structured fields: every important value should retain enough provenance to trace it back to the page, region, table, row or passage that produced it.
Why it matters
AI can turn PDFs, scans, spreadsheets and mixed documents into convenient JSON, but convenience becomes dangerous when extracted values lose their source context. A number can be copied from the wrong column, a footnote can change its meaning, or OCR can silently alter a unit. Downstream analysis is only defensible when extraction preserves evidence and exposes uncertainty instead of presenting every field as equally reliable.
01
Separate capture, normalization and validation
First capture source-local evidence such as text blocks, table cells, coordinates or page references. Normalize that evidence into the target schema without discarding the original representation. Then validate types, units, required fields and cross-field constraints. When extraction is ambiguous, keep candidate values or an explicit low-confidence state and route the field to review instead of fabricating certainty.
02
Example: extracting quarterly revenue
A report contains a revenue table with values in millions and a footnote excluding one business unit. The extractor stores the numeric value together with page, table, row, column, unit and footnote reference. A later analysis can therefore distinguish a genuine revenue change from a parsing mistake or a changed reporting definition.
Common failure modes
- Flattening extracted values into JSON with no source location or provenance.
- Normalizing units or dates silently without retaining the original value.
- Forcing uncertain cells into a single answer instead of preserving ambiguity.
Engineering heuristics
- Attach source coordinates or stable citations to consequential extracted fields.
- Validate units, ranges and relationships after extraction, not only syntax.
- Keep raw evidence available so disputed values can be re-inspected.
Takeaways
- 01Extraction quality includes provenance, not only field accuracy.
- 02Normalization should not destroy the path back to the source.
- 03Uncertain extraction should remain explicitly uncertain until verified.
Related concepts from the Knowledge Graph
These relationships come from the canonical graph, not a separate Guide taxonomy.