This is the normal case during ramp-up and for every new document type — and this is exactly where a classic chain ends: A trained extraction model requires labeled sample documents, a training run, and waiting time. NexoGraph Document Processing does not need that: there, an analyzer is authorized instead of trained — a field schema that is published is enough. There are two ways to create this schema, and the difference determines whether the chain stops here or continues.
edit_square
Path A · manual
Describe the schema in the editor
A subject-matter admin defines field names, data types, descriptions, and table columns in the configuration console and publishes the analyzer. Full control over the business semantics — but a human step.
Document waits until maintained
auto_awesome
Path B · AI-generated
Have the schema suggested from the document
The AI reads the triggering document and derives the field schema from it — fields, types, descriptions, tables. No person, no appointment, no waiting time: the model is created in the current run.
Document continues without interruption
Only Path B fulfills the promise of the home page. With the manual path, the chain is partially automated for an unknown document type — it stops and waits for configuration. Only the AI suggestion makes it fully automated: the system expands its own configuration state while it works. The subject-matter admin then reviews and refines the schema — not before.
Path B in the current run
STEP 1
AI suggests schema
The triggering document is analyzed, and the field schema is derived from it — field names, types, descriptions.
STEP 2
Publish analyzer
The schema is published as an analyzer — no training run, immediately usable, idempotent.
STEP 3
Attach example
The document remains attached to the model as a permanent example — later, the subject-matter admin finds a prepared schema editor.
STEP 4
Continue processing directly
The same run analyzes the document with the new model — no second attempt, no waiting time.
lockDelivery remains locked. The assignment to the target system is deliberately not created with it — an automatically created model delivers data to the review list, never unchecked into the target system.
filter_altNo cost avalanche. It is triggered only for categories the classifier has actually learned, and at most once per category — every further case uses the model that was created.
toggle_offCan be switched off, with a defined fallback. If the function is off — or if a parallel run was faster — the case is skipped in the journal with a plain-language reason. A document disappears in neither branch.
tuneTwo engines, one result format. For stable mass documents, the trained model remains the more accurate choice; the authorized schema is the path for new and rare cases. Both deliver the same result format —
final processing does not notice any difference.
