Changelog
Product releases affecting the API, the pipeline engine and the review surfaces.
Parse, Split, Classify and Extract documented in depth
New reference pages break each Parse-side stage down to its sub-stages — the Parse phase (layout, language, OCR), Split, Classify, and Extract — with the confidence and routing behaviour each one carries.
Accounts and billing docs
Added guides for the three account surfaces: generating a scoped API key, how credits are metered (reserve, settle, release) and how to top up, and bringing your own LLM key so inference runs on your provider account.
Stage errors split from exceptions
Hard stage failures now surface on their own error-monitoring view in Jobs rather than mixing into review queues. A document that merely needs a human is unaffected.
Documents API
GET /v1/documents ships with lifecycle attributes, lineage and a per-document event ledger. Previously only reachable through the parent job.
Per-operator coverage rows
Verification coverage moved from a percentage on the rule to a flat list of rows on the person. A correction's coverage is the highest percentage among matching rows.
Field-value rule conditions
Coverage rules can now key off an extracted field value — total ≥ 50000, class = credit_memo — in addition to operator and exception attributes.
sample_rate, the ramp fields and priority were removed from review routing rules; a rule is now a pure matcher. Set rates on per-person coverage rows instead.
Taxonomy creation from a folder upload
Upload a folder tree and IDPForge derives classes, codes and sample documents from it — folders become classes, contents become few-shot examples.
Pipeline versions are immutable
Editing a published pipeline now creates draft edits; publishing mints the next version. In-flight jobs finish on the version they started with, and every job records its version.