Training tab (second full tab: corpus, accuracy, correction loop)¶
| Stage | Built |
|---|---|
| Status | Shipped |
| Design status | In review |
| Estimate | M |
| Linear | PIN-684 ↗ PIN-682 ↗ |
| Linear status | Done |
| Cycle | C14-C15 |
| Design | Design ↗ |
| Linear epic | Data Ingestion & Integrations |
| Module | Smart Import ↗ |
Priority: Medium
Scope¶
Captures manual corrections made while staging a Smart Import batch — folder moves, vendor equipment/installation tags, CO-naming fixes — and feeds them back into AI classification so later imports get more accurate on the same patterns. Delivered as the staging review UI (reorder/rename/reclassify/reassign before approve) plus a correction-capture and feedback loop; these two tickets do not ship a separate corpus/accuracy analytics tab. (Source: PIN-684, PIN-682.)
Acceptance criteria¶
- Staged files list an AI-proposed folder/module/project assignment with a confidence badge; low-confidence items surface first
- User can reorder, rename, reclassify, and reassign staged files before approving
- Approve hands the batch to commit, and captures each correction (original AI proposal vs. user-corrected value) for training
- Corrections persist and feed back into classification (few-shot / correction lookup) on the next import
- First-pass classification accuracy measurably improves on previously-corrected patterns
- Loop degrades gracefully (non-fatal) when no correction history exists yet
Conor's comments¶
September
Notes¶
Checked all 110 rows: ROW 22/23/24 are SharePoint/Excel ingest, ROW 1 is backend model-training log capture, ROW 36 is Cara chat feedback/digest - none is an accuracy/corpus analytics tab, and marking. Linear (PIN-684, B6 correction-learning loop) shows Done — flagging for review since spec status is deferred/backlog; confirm whether a distinct Training-tab UI has shipped or only the backend loop (PIN-682, B4 staging-review UI, also Done, is the closest UI-facing sibling).