Documents read more accurately

Four weeks of accuracy work on the extraction pipeline. Most of this you’ll only notice as the absence of something annoying.

Classification

• The filename now goes into the classification prompt as a weighted hint, with a glossary of the abbreviations exporters actually use — FCT/FAC/FE for invoices, PL for packing lists, BL/AWB for transport documents, FITO for phytosanitary certificates, EUR1/CO for certificates of origin. Previously the AI classified blind to the filename.
• Courier waybills (DHL, FedEx, UPS) are a document type in their own right, so they no longer land in the wrong bucket.
• Admins can recategorize a document from a dropdown next to its type badge. Trackberry re-extracts it as the chosen type and cleans up whatever the old type imported.

Spreadsheets and PDFs

• Hidden columns in Excel packing lists are properly excluded — visible columns win when the same header appears twice.
• More Spanish headers recognised: Calidad as category, Envase as format/box weight, abbreviated Calib. as calibre. Category values shed their “Category”/“Cat”/“Categoría” prefixes.
• A padded summary sheet no longer wins over the detail sheet.
• A packing list whose header row isn’t ruled off gets its column names rebuilt from the header’s own geometry, instead of collapsing 18 labels into one cell and leaving the AI to guess which number was the calibre.
• Alphanumeric calibres (1X, 2X) survive extraction intact.

Counts, weights and pallets

• Box counts reconcile against the declared totals and fail loudly rather than quietly doubling.
• Missing box weights are derived from net weight ÷ box count, and asparagus “atados” from tie count × weight per tie.
• Pallets sort naturally by number instead of lexicographically, so 10 comes after 9.

References

• Hong Kong phone numbers are structurally identical to IATA air waybill numbers, and one had been filed as a shipment’s AWB. Candidate numbers now have to pass the IATA check-digit test, and a valid AWB from an attached waybill beats one scraped out of an email signature.
• Duplicate shipments from separator differences in AWB numbers (“160-14832020” vs “16014832020”) are matched as the same reference.

Alongside this, the Needs Assignment picker is now searchable and grouped into Sea and Air, with each shipment labelled by its container or transport reference rather than an internal SHIP- code.

If a document still reads wrong, Troubleshooting covers what to do about it.

FAQ #

Do I need to re-upload documents that were read wrong before this? #

Only the ones that matter. An admin can recategorize a document from the dropdown next to its type badge, which re-extracts it and cleans up whatever the previous type imported.

Why does a box count now fail a check instead of being corrected? #

Because a count that disagrees with the declared total means one of the two readings is wrong, and guessing which would occasionally double a shipment silently. Failing loudly puts it in front of a person instead.

Does Trackberry only read English documents? #

No. Spanish packing list headers are recognised directly, and free-text values in any language are understood — the extraction works from what a document means rather than from a fixed template.