Overview
Comments
Database
KPIs
Infra
Period:
2 days
Overall Accuracy
Overall Accuracy
--%
Loading...
Accuracy w/o Ref & Repère
--%
sans référence & repère
Perfect Extractions
--%
validés SANS modification
Perfect w/o Ref & Repère
--%
hors référence/repère
Perfect Line Rate
--%
par ligne (measurement)
Perfect Line w/o Ref & Repère
--%
lignes hors ref/repère
Extractions
--
verified in period
Lines (Measurements)
--
total measurements analyzed
High Confidence (>=80%)
--%
Monce conf >= 80%
Verification Rate
--%
verified / total
Article Matching Accuracy
Dimensions
Quantity
--%
AI extraction vs human correction
Largeur
--%
AI extraction vs human correction
Hauteur
--%
AI extraction vs human correction
Text Extraction & Client Matching
Reference
--%
AI extraction vs human correction
Repere
--%
AI extraction vs human correction
Client
--%
client_matching.nom vs corrected
Accuracy by Factory
Factory
Overall
Verre 1
Verre 2
Intercalaire
Remplissage
Faconnage
QTY
Largeur
Hauteur
Verified
Loading...
Accuracy by Client
Client
Overall
Verre 1
Verre 2
Intercalaire
Faconnage
QTY
Largeur
Hauteur
Client Match
Verified
Loading...
Examples: What Works vs What Doesn't
Verre 1
Verre 2
Intercalaire
Remplissage
Faconnage
Client
All Factories
Correct Matches
Loading...
Show All
Incorrect Matches
Loading...
Show All
Monce Expression (Snake Match Types)
How Monce matches queries: Exact (100% synonym match) vs SAT (>=50% lookalike vote) vs Low Conf (<50%)
Field
Exact %
SAT %
Low Conf %
Total
Loading...
Daily Volume by Factory
All Factories
VIT
VIP
Monce
Synonym Management
Apply Approved Synonyms
Batch-apply all approved synonyms to Monce models. Duplicates (100% matches) will be skipped.
0 approved
Rebuild Monce Models
Unmatched queries from verified extractions. Approve suggestions to add them in the next rebuild.
Article Suggestions
Loading...
Text
Target Article
Count
Actions
Loading...
Client Suggestions
Loading...
Text
Target Client
Count
Actions
Loading...
Methodology
Data Sources
Production RDS - Aurora MySQL (monce-ai-prod-db)
Table - extracted_data avec status = 'verified'
Ground Truth - Monce Matching API (SAT-based article matcher)
How Accuracy is Calculated
Fetch - Extractions verifiees des N derniers jours avec value_corrected
Extract - Texte corrige par l'humain: value_corrected.measurements[i].verre1, etc.
Ground Truth - Appel batch a Monce API pour obtenir le num_article correct
Compare - Prediction AI (matching[i].verre1.num_article) vs Monce result
Formula
Accuracy = (AI predictions matching Monce) / (Total comparisons) x 100
Fields Analyzed
verre1 Article principal (le plus difficile - "black holes" Monce)
verre2 Deuxieme vitrage (double/triple vitrage)
intercalaire Espaceur entre vitrages (Warm-edge, Alu...)
remplissage Gaz (Argon, Air, Krypton)
Perfect Rate Metrics
Perfect Extraction Rate % of documents (orders) where ALL fields across ALL lines are correct. Strict: a 20-line order with 1 bad field fails entirely.
Perfect Line Rate % of individual measurements (lines) where all evaluated fields are correct. More granular: a failed document can still have 95% perfect lines.
Perfect Line w/o Ref & Repère Same as Perfect Line Rate but only evaluating article fields (verre1, verre2, intercalaire, remplissage, faconnage). Excludes reference and repère text fields.
Perfect Line Rate = (lines with all fields correct) / (total lines with data) x 100
Invariant: Perfect Line Rate >= Perfect Extraction Rate (always)
Known Limitations
verre1 ~44% - Monce "black holes" (#60442, #1004) absorbent les variantes (LowE, ITR, SATINOVO)
Missing data - Champs vides dans value_corrected = pas de comparaison possible
Monce = Ground Truth - Si Monce se trompe, on compte comme erreur AI meme si AI avait raison