Document classificationwith chained classifiers.
Every document is checked against types you define, by a chain of classifiers that starts cheap and escalates only when needed: filename rules, structure checks, then semantic evaluation by an LLM.
A chain, not a black box.
Ordered classifiers run until one matches; you control the order and the rules.
Filename rules
Regex and pattern checks catch the documents your senders already label, at zero model cost.
Structure checks
Single- and multi-structure classifiers test whether the document actually contains what its type requires.
Semantic evaluation
An LLM evaluates your written condition for the cases rules cannot separate, including an exclusive mode that picks exactly one type for mixed inboxes.
Your types, not a vendor catalog.
Pre-trained classification ties you to the vendor's idea of what documents exist. Chained classification evaluates your definitions, so a document class specific to your firm classifies as reliably as an invoice.
The chain also gives you an economics dial: cheap deterministic checks first, model-based evaluation only for what remains, and a review queue for what no classifier places confidently. In mixed-inbox pipelines, the exclusive chooser mode assigns the matching schema automatically, so one linear pipeline handles many document types.
Classification questions
What teams designing classification ask most.
Classifiers run in the order you configure until one matches: filename rules first, then structure-based checks, then semantic LLM evaluation. Each classifier is a config you can edit and reorder, so the decision logic is inspectable rather than a single opaque model score.
Yes: the multi-structure classifier's exclusive mode makes one chooser call that picks exactly one of your types (or none), and the extraction stage then applies that type's schema automatically. One linear pipeline processes invoices, contracts and forms arriving in the same mailbox.
That is the point of user-defined types: a classifier chain evaluates your definitions and conditions, not a fixed catalog. If you can describe the type (by name patterns, expected structure, or a semantic condition in plain language), the chain can classify it, and the review queue catches what remains ambiguous.
See it on your own documents.
The 14-day free trial covers 200 documents: enough to define your schemas, run real files and measure the results before anyone calls you.