POA Interpretation Workflow
Summary
POA interpretation converts source documents into typed procurement evidence: document kind, header fields, line items, vendor hints, job hints, PO hints, invoice data, confirmation data, credit memo data, and packing slip evidence. It is not allowed to be the final authority; it creates structured claims that matching, validation, repair, and sync can inspect.
The concept page POA Interpretation covers the primitive. This thread frames interpretation as one stage in a larger operator-repairable workflow.
Reader Question
How does POA turn a document into structured evidence while preserving enough uncertainty for later validation and repair?
Surface Or Workflow
Operators see interpretation output as document previews, fields, lines, diagnostic callouts, and repairable values. Engineers enter through document runtime orchestration, per-kind processors, interpretation contracts, validation, LLM/OCR cache behavior, and persistence repositories.
Lifecycle
After intake, document runtime resolves the source, creates an attempt, chooses the appropriate processor, and persists interpretation artifacts. Per-document processors handle invoices, packing slips, confirmations, credit memos, and shared extraction behavior.
The interpretation layer validates shape and confidence before downstream matching. If a document kind is ambiguous or a field cannot be trusted, the workflow records that condition for review instead of pretending the extracted text is canonical.
Child Threads
- POA Matching And Validation Workflow: decides whether interpreted evidence matches operational records.
- POA Operator Repair: exposes document-kind and field repairs.
- AI In Operational Workflows: explains why model output remains evidence, not authority.
Implementation Boundaries
services/poa/document_runtime owns orchestration and processors. services/poa/interpretation owns interpretation contracts and validation. services/poa/persistence/interpretations stores the result. The thin case processing path reaches this behavior through services/poa/jobs and jobs/poa/process_case/run.py.
Tradeoffs
Keeping interpretation separate from matching means the same extracted evidence can be inspected, repaired, or reused. The cost is more intermediate state, but that state is what lets POA explain why an invoice failed for interpretation versus why it failed for PO matching or external command eligibility.
Visual
The current visual is a graph neighborhood. The intended visual is an interpretation sequence from source artifact to processor, normalized interpretation, validation signals, persisted artifact, and downstream matching.
Source Evidence
docs/engineering/concepts/poa-interpretation.mdservices/poa/document_runtimeservices/poa/document_runtime/processorsservices/poa/interpretationservices/poa/persistence/interpretationsservices/poa/jobs/case_interpretation.pytests/poa/test_poa_interpretation_pipeline.pytests/poa/test_poa_interpretation_persistence.py
Related Threads
- POA System
- OCR And LLM Services
- AI In Operational Workflows