cadrumo.application.ledger.invoice_draft_extraction module¶

On-host routing of one stored evidence reference to a typed invoice draft.

Holds the reading layer that decides which reader a given document is handed to, and the three readers’ wiring. It never persists an Invoice and never guesses a value no reader could ground in the document: every field a reader could not recover is left None rather than fabricated (no-silent-under-declaration in spirit: an unconfident field is absent, not invented).

Three reading paths, chosen on the document’s own shape rather than on its stored MIME type:

A structured e-invoice (Facturae, CII, UBL) is read exactly through the application-owned StructuredInvoiceReader. It reaches no model at all, so prompt injection is categorically impossible for that document rather than merely mitigated, and it is the only path that can recover the document’s own line decomposition and per-rate breakdown.
A text-native PDF is transcribed by transcribe_text_layer() and read first by the fixed label rules of read_invoice_fields_by_labels(). When those rules read every required field no model runs; otherwise the fields they could not read come from extract_invoice_fields_from_text(), and an unreachable model leaves the rule reading standing. Either way the draft is grounded against that same transcription by ground_draft_against_transcription(). The transcription is produced by a DIFFERENT reader than the one that proposes values, which is what makes the anchor check an external check rather than a model confirming itself.
A scan-only PDF or image has nothing to transcribe, so it falls back to the on-host LOCAL vision reader (evidence_draft_vision) – the same rasterise-then-read-with-Ollama transport LocalVisionLLMClassifier already uses for classification, gated by LLM_VISION and never a cloud call.

The escalation is one-directional and the asymmetry is deliberate. A document with no text layer genuinely needs a reader that works on pixels, so it escalates. A readable document whose semantic reader is absent or unreachable is an ENVIRONMENT failure, so it REFUSES in the operator’s face rather than silently running a heavier engine they did not ask for – see _refuse_a_text_read_with_no_reader(). A profile that has opted out of vision reading likewise gets a typed, instructive refusal, never a silent empty draft.

Everything here runs on-host and in-memory only. The evidence bytes, the transcription and the draft never touch disk and are never sent to a cloud provider. This module performs no filesystem write.

extract_invoice_draft_from_evidence() is the CLI-facing wiring layer: it resolves an already-stored purchase_invoice_evidence record or a linked attachment_id to its in-memory bytes (through the private evidence-input resolvers resolve_purchase_invoice_evidence_input() and resolve_attachment_evidence_input()) and routes them to one of the three readers above, so aeat app ledger evidence extract needs only a bucket id plus one of the two reference ids.

Confirming a read draft into a catalogue record is a separate responsibility and lives in invoice_confirmation; the draft’s own direction-resolved projections live in evidence_draft.