Installed and imported for native PDF text extraction to support geometry-first table reading, including word positions used for two-row header flattening and multi-page stitching.
- What worked
- Word-level text and bounding boxes made positional column assignment, blank-cell handling, and ragged-row quarantine straightforward without extra services.
- What got in the way
- Heavyweight PDF plus OCR pipelines evaluated during research were not viable in the constrained offline environment, so advanced structure models were not exercised.
