Documents — one table, five states

Processing — in flight

Documents mid-read. Confidence reads 0% because there is nothing to score yet, and the page count is already known. This filter should always be nearly empty; if it is holding steady at a large number, documents are stuck rather than processing.

Failed — look for the shape, not the file

The valuable habit on this filter is looking for repetition. In the filmed week the failures were the same shape of file over and over — a spreadsheet from one sender, 0% confidence every time. That is one fix (a carrier profile, or a word with the sender about format), not eight investigations. Failures almost always cluster.

Needs Review — read cleanly, waiting on a person

The counter-intuitive filter. In the filmed tenant, Needs Review held 453 pages of invoices at 100% confidence. High confidence does not mean auto-approved — a document lands here because a business rule says a human signs off, not because the read was doubtful. This is a queue about policy, not accuracy.

• Not an error — 100% confidence and Needs Review coexist perfectly happily.
• Work it elsewhere — The Review and Validation screens are built for clearing this.
• Watch the size — A growing Needs Review pile is a staffing signal, not a quality one.

Extracted — the good case

A named carrier, a high confidence score, and status Extracted. This is the state everything is trying to reach. Beside the status filter sits a carrier multi-select, which is how you answer the useful version of the question: not "how are we doing?" but "how are we doing on this carrier?" Filter to one carrier and the quality of its extraction profile is immediately legible.

Uploading a document by hand

Upload New Document opens a drawer for the documents that never arrived by email — something handed over at a dock, downloaded from a carrier portal, or re-sent after a bounce. The drawer includes a carrier picker, and choosing the right carrier there matters: it is what tells KnowDoc which extraction profile to read the file with.

What the five filters say together

Read in order they form a diagnosis. Processing should be near zero. Failed should be small and, when it isn't, should be one repeated cause. Needs Review should be steady rather than growing. Extracted should be most of the volume. Any filter that breaks its own pattern points at a specific screen: Carriers for extraction problems, Flagged for exceptions, Review and Validation for the human backlog.