Dashboard
| Project | Jobs | Configs | Docs | Coverage | Cached | Tag triage (G / A / R) | DI $ | GPT $ | Total $ |
|---|---|---|---|---|---|---|---|---|---|
| KRL-P-700 | 2 | 2 | 12/12 |
100%
|
100% |
|
β | β | β |
Costs are indicative β list-price estimates from recorded usage (DI pages Γ price + GPT tokens Γ price), apportioned to projects by document share. A guide for relative comparison, not the Azure invoice; covers runs recorded since per-job cost tracking was added.
π― Measured accuracy (golden set)
Upload a hand-checked answer file (JSON: {"DOC-NUMBER": ["TAG-1","TAG-2"], β¦}).
Scores what Lattice extracted against your answers and reports precision, recall and F1 β the authoritative
accuracy number. Documents are matched by number, filename or stem.
Coverage is exact (cached results counted). Avoided cost is an estimate β cached docs Γ average cost per fresh document β framed per re-run, not a running tally of money already saved.
βΆ Sample of register tags never found (25 of 33)
Reconciled against the currently-loaded register across all documents. Extra includes real tags absent from the register as well as any unrecognised strings.
GPT silently samples dense tables β it returns clean output but omits scattered rows. After GPT returns, a deterministic regex sweep of the full document text adds back any tag it dropped, flagged with provenance. This figure is that safety net's contribution: the tags that would otherwise have been silently lost. Counted from the latest cached result per document.
Cost trend appears once runs with recorded cost exist (per-job cost tracking is recent).
| Month | Jobs | Docs | DI $ | GPT $ | Total $ |
|---|---|---|---|---|---|
| 2026-10 | 2 | 24 | β | β | β |
Bucketed by completion month, last 12 active months. Cost is indicative (list-price estimate), and only runs recorded since per-job cost tracking was added carry a figure.
| Config | Jobs | Docs | Avg tags/doc | Triage (G / A / R) | Unmatched | Cost | Cost/doc |
|---|---|---|---|---|---|---|---|
| Demo: tags, vendor, PO, size and material | 2 | 12 | 52.2 |
|
0.2% | β | β |
Yield and triage are computed from each config's cached results (the same G/A/R logic as the rest of the dashboard). Unmatched is the share of a config's tags that match neither a tag-class pattern nor the register. Cost is indicative and attributed from the run log; configs run before per-job cost tracking show no cost.
| Project | Missing PDFs | No tags | To normalise | Unmatched |
|---|---|---|---|---|
| KRL-P-700 | 0 | 0 | 0 | 42 |